WO2020147920A1

WO2020147920A1 - Traffic signal control by spatio-temporal extended search space of traffic states

Info

Publication number: WO2020147920A1
Application number: PCT/EP2019/050828
Authority: WO
Inventors: Radu TUDORAN; Stefano BORTOLI; Cristian AXENIE; Mohamad Al Hajj HASSAN; Goetz BRASCHE; Hailin Li
Original assignee: Huawei Technologies Co., Ltd.
Priority date: 2019-01-14
Filing date: 2019-01-14
Publication date: 2020-07-23
Also published as: CN113316808A; CN113316808B

Abstract

The present disclosure relates to an apparatus and method for traffic control of vehicles at one or more intersections via traffic signals. Each intersection may have a plurality of lanes and the lane may include a traffic signal and a traffic signal timing, defining a state of a lane. For each of the plurality of states of a first intersection, traffic data is acquired and may include traffic data from lanes of intersections neighboring the first intersection. The acquired traffic data are used to predict traffic data at multiple predefined timings of traffic signals for each state. Among the predefined timings, a traffic signal timing is determined for each lane by optimizing a predetermined function of the predicted traffic data of all lanes. The traffic signal timings are then used to control the traffic signals for the one or more lanes.

Description

TRAFFIC SIGNAL CONTROL BY SPATIO-TEMPORAL EXTENDED SEARCH

SPACE OF TRAFFIC STATES

TECHNICAL FIELD

The present disclosure relates to traffic control using optimized traffic signal sequences.

BACKGROUND

Traffic congestion is becoming major issues in cities and metropolitan areas in most countries. Various factors have been identified that contribute to traffic congestion, such as bad road conditions, inefficient traffic flow controls, increased number of vehicles on the roads, and even at times ill-mannered practices of road users. With increasing population and hence traffic volume, traffic congestion poses serious challenges toward the city infrastructure facilities and also affect the socio economic lives of the people due to time wasted while waiting in traffic. Statistics show that the average annual traffic congestion cost in the United States in 2014 was 1433 dollars per auto commuter, or over 5 billion dollars per city for very large urban areas.

In general, traffic includes the flow of vehicles such as cars, vans, motorcycles or the like on streets/roads in cities, rural areas, or (interstate) highways, all of which contribute to the overall flow of the participating vehicles. Moreover, the traffic on one road may be directly or indirectly impacted by the traffic on other neighboring roads and may depend also on the particular time of the traffic, for example at morning or evening rush hours. As such, traffic has been abstracted through so-called traffic-flows, which are a complex spatial and temporal process as result of the driving behavior the traffic parties. A variety of characteristics can be distinguished in traffic- flows, such as volume, average speed and speed distributions, headways and travel times and the like. The formal-mathematical description of the relationship between these differing traffic- flow characteristics is known as "traffic-flow" models.

While various solutions for optimizing the traffic based on traffic flow models have been developed in order to tackle this problem, none of the solution yet excels in terms of the capability to generalize and adapt to new situations, the precision in predicting traffic, fast reaction times (real-time processing of traffic data) and system complexity. SUMMARY

Embodiments of the invention are defined by the features of the independent claims, and further advantageous implementations of the embodiments by the features of the dependent claims.

In particular, the present disclosure targets the area of traffic control systems enhanced to optimize the traffic by leveraging a real-time engine that applies Statistical and Machine Learning for Big Data Distributed Stream Processing. Traffic metrics are received as streams of traffic data, which are sequences of events (i.e. tuples containing various types of data based on a traffic metric, such as number of cars, speed of cars etc.) that are collected from various sources (e.g., cars, sensors: traffic light cameras, street induction sensors etc.) in a chronologically ordered fashion. The stream processing paradigm involves applying business analytics, or more complex learning functions over the events in the stream, for example predicting the traffic flow over time on a road. A typical approach to stream processing assumes accumulating such events within certain boundaries at a given time and applying business analytics functions on the resulting collection.

According to an aspect of the present disclosure, an apparatus is provided for traffic signal control at a first intersection including a plurality of lanes and a traffic signal for each of the plurality of lanes. The apparatus comprises a processing circuitry configured to acquire traffic data for a state among a plurality of states, wherein the state is a combination of one or more lanes belonging to the first intersection and a same traffic signal timing of the traffic signal for the one or more lanes; predict traffic data for each state of the plurality of states based on the acquired traffic data; determine for each state among the plurality of states a first traffic signal timing out of a plurality of predetermined traffic signal timings by optimizing a predetermined function of the predicted traffic data; and control a first traffic signal for each of the plurality of lanes according to the determined first traffic signal timing.

The use of a state description (state-space approach) with a state including one or more lanes and a same traffic signal timing for the one or more lanes provides the advantage that the number of states that have to be managed for traffic flow prediction and optimization is reduced. Hence, the state reduction performed in this manner ultimately reduces the amount of data to be stored and/or need to be held available in order to perform traffic flow optimization in real time. Simultaneously, the cost spent on hardware (e.g., for storing the data) may be reduced, since the traffic data are acquired for a reduced set of states. As a result, the prediction of traffic flow may be performed faster since the data acquisition is based on a reduced state space.

According to an embodiment, the traffic data is based on a traffic metric with the traffic metric being any one or more of a number of vehicles on the lane, an average speed of vehicles on the lane, a vehicle occupation of the lane, a vehicle queue length of the lane, and/or an average waiting time of vehicles on a lane.

This means that the approach of the disclosure is able to handle a variety of different metrics alone and/or in a combined manner. This provides the advantage that the traffic prediction may be adapted by choosing one or more metrics that is most suited (alone or combined with other metrics) for a particular intersection, depending on the specific environment of said intersection.

According to another aspect of the present disclosure, the processing circuitry of the apparatus is further configured to optimize the traffic data according to a constraint, including one or more of a maximum cycle time of the traffic signal timing, minimum green time per lane, yellow-light time, a constraint on that there is always a green light for turning right, or the like.

According to an example, the constraint is predefined or selected based on the traffic metric.

Hence, the approach of the present disclosure enables a constraint optimization, accounting for time constraints for example. This restricts further the search space for the states, having timings that optimize the traffic flow. This may accelerate the traffic prediction even further since an optimal solution is found in a lower dimensional state space.

According to an embodiment of the present disclosure, the plurality of states includes states of a plurality of lanes of a second intersection spatially neighboring the first intersection.

For example, the acquiring of the traffic data for a state includes traffic data of a state of a lane among the plurality of lanes of the second intersection.

While the prediction of traffic flow for lanes of an intersection may be view as a local operation, the local traffic however is usually impacted by the traffic flow at other neighboring intersections. This may depend also on the mutual distance between intersections and the region where intersections are located. Hence, accounting for states of spatially neighboring intersections provides the advantage of accounting for spatial correlations on different spatial length scales, as inherent to traffic flow behavior (e.g., cooperative behavior of traffic members (vehicles etc.). This improves further the accuracy of the traffic prediction, including possible different environments of the intersection.

According to an aspect of the present disclosure, the traffic data acquisition is further based on a traffic data history and the processing circuitry is configured to acquire traffic data from the traffic data history acquired at a plurality of time points earlier than a time point of the traffic data acquisition. Traffic data carry inherently not only spatial correlations, but also temporal correlations. This means that in order to make a prediction for a traffic flow (i.e. into the temporal future), the evolution of traffic prior to the time of the prediction may be utilized. Hence, the use of traffic data history allows to account for traffic characteristics in the past, which may vary for example for day time and evening times, rush hour or the like. Moreover, the past temporal behavior may also depend on the location of the intersection. Therefore, traffic data history enables to improved further the accuracy of the traffic prediction by including temporal correlations.

In summary, the approach of the present disclosure is able to prediction traffic flow for an intersection with high accuracy by utilizing spatial-temporal correlations of the traffic data. This may include correlations in different spatial and temporal length scales (i.e. the degree of the non-Markovian characteristics of the traffic data in space and time).

According to a further aspect of the present disclosure, the prediction of the traffic data is based on a parametric prediction model for the state and the processing circuitry is configured to update a parameter of the prediction model in accordance to the acquired traffic data; the predicted traffic data; and/or the spatially neighboring second intersection.

Traffic flow data representing the “behavior” of vehicles (manned and/or un-manned / autonomous) through a traffic metric are inherently dynamic, which makes the prediction of traffic challenging. Part of the dynamic traffic characteristics is accounted for the in traffic prediction by acquiring the traffic data as used for the prediction at a current / latest time and / or in conjunction with data acquired less currently (past data according to a traffic history).

The inherent dynamic nature of traffic data may be extended further in the approach by use of a predictor model for the traffic flow that is itself dynamic. This means that the model itself is adaptive, i.e. the one or more model parameters are updated. Updating the parameters in accordance to acquired traffic data (i.e. most recent), predicted traffic data, and/or spatially neighboring one or more intersections, the predictor model with its representing parameters is permanently updated. In other words, the model itself is up-to-date and adjusted to the current spatial-temporal dynamics of the traffic flow.

As a result, the traffic prediction is performed with high precision, since the predictor parameters are customized with reference to latest traffic data. Moreover, since the approach of the present disclosure is based on adaptation of the predictor, the traffic flow predictions system may be easily matched and adjusted to many different traffic environments, including cities and areas with high or low intersection densities and / or roads having a different number of multiple lanes. For example, the processing circuitry of the apparatus is further configured to update the parameter of the prediction model of a first state based on a second state selected among the plurality of states.

According to an aspect of the present disclosure, the selection of the second state is based on a selection policy referring to a state for which traffic data has been acquired, a state for which traffic data has been predicted, and/or a state whose traffic data has been obtained by transforming a traffic metric of a different state.

The above-mentioned adaption may be fine-tuned by updating the parameter based on other (i.e. second) states. In addition, the selection itself may be diversified for a state by performing the selection dependent on certain policies or the like, including states for which traffic data has been acquired and/or predicted. The selection may even be based on transforming the traffic metric of a different state. In other words, different traffic metrics of different states may be linked with each other. This means of parameter update on a state-basis therefore enhances the flexibility of the traffic flow predictor system even further.

According to an aspect of the present disclosure, a method is provided for traffic signal control at a first intersection including a plurality of lanes and a traffic signal for each of the plurality of lanes, comprising the steps of acquiring traffic data for a state among a plurality of states, wherein the state is a combination of one or more lanes belonging to the first intersection and the traffic signal for the one or more lanes being same over a traffic signal timing out of a plurality of predetermined traffic signal timings; predicting traffic data for each state of the plurality of states based on the acquired traffic data; determining for each state among the plurality of states a first traffic signal timing out of the plurality of predetermined traffic signal timings by optimizing a predetermined function of the predicted traffic data; and controlling a first traffic signal for each of the plurality of lanes according to the determined first traffic signal timing.

According to an aspect of the present disclosure, a computer-readable non-transitory medium is provided for storing a program, including instructions which when executed on a processor cause the processor to perform the steps of the method.

Hence, the present disclosure introduces a new real-time processing system to: collect traffic metrics, model them in search states, make a prediction about the future traffic flow and select from the search states the signal control sequence to enhance the road traffic flow. Therefore, the optimization of the traffic is done continuously and in real-time from an incoming stream of traffic metrics (e.g., number of cars, speed of cars, occupation at traffic light, etc.). The considered scenario is thus well-suited for such continuous, real-time learning and adaptation (i.e. with very low latencies with respect to the time reference of the most recent update - last incoming event). Moreover, in traffic flow prediction and control, the control unit must estimate and accommodate changes in the stream data distribution and provide accurate predictions and judicious control actions (i.e. traffic light green color timings) despite the single pass over the incoming data. As the data stream progresses with the traffic evolution, the computations have a limited time span to be handled in the system, thus ensuring a bounded resource allocation and execution time.

The present disclosure overcomes the resource greedy, computationally expensive and complex state-of-the-art approaches (e.g., complex analytical flow models based differential equations and numerical methods, empirical methods, neural networks), by the new specialized stream traffic control compute unit, that exploits the spatial and temporal correlations among the different traffic metrics describing the traffic situation. The control unit output can be applied to traffic lights in order to maximize the traffic flow. The proposed unit can work on any traffic metric based on which will output a time configuration to be used by the next traffic light cycle. The system is supported by a flexible instrumentation ensuring updates with low-latency, high incoming event rates and a fixed resource budget. Further, the system can be deployed to any type of intersection without pre-training, which offers major advantages in terms of deployment costs reduction.

Details of one or more embodiments are set forth in the accompanying drawings and the description below. Other features, objects, and advantages will be apparent from the description, drawings, and claims.

BRIEF DESCRIPTION OF THE DRAWINGS

In the following, embodiments of the invention are described in more detail with reference to the attached figures and drawings, in which:

Fig. 1 is a sample intersection comprising four lanes, with their traffic light timings being the same to control the traffic flow.

Fig. 2 is a sample intersection comprising four lanes, with their traffic light timings being different to control the traffic flow.

Fig. 3 is a sketch of a window aggregator memory, getting out of memory space during continuous traffic data acquisition. Fig. 4 illustrates the use of traffic data acquired for a four-lane intersection, provided as input to a neural network performing the prediction of the traffic flow, along with the traffic light timings at the output layer of the network.

Fig. 5 is a schematic drawing of an embodiment, with data acquired within a finite-sized window and used to perform prediction and control of traffic lights in real-time.

Fig. 6 is a block diagram of an embodiment, including a traffic predictor, a traffic optimizer, and a predictor updater. Traffic data is acquired for a state and new data is predicted by functional optimizing under constraints. The predictor model is adaptive, with its parameters being updated based on acquired and/or predicted data, including data from states of neighboring intersections.

Fig. 7 is an overview of an architecture of an online traffic controller according to an embodiment.

Fig. 8 is a functional architecture of traffic light optimization, including two prediction models of different states belonging to the same intersection, exchanging model parameters. Fig. 9 is an illustration of a traffic light optimization problem, including multiple lanes reduced to a smaller number of states (state reduction).

Fig. 10 is an illustration of the traffic light optimization problem, including a selection of states with possible varying timings so as to optimize the traffic flow due to a traffic metric.

Fig. 1 1 is an illustration of the traffic light optimization problem, including an update of the predictor model to perform traffic prediction for the next time step.

Fig. 12 is an overview of an online traffic control system for one intersection according to an embodiment.

Fig. 13 is an overview of an online traffic control system for a sample city including multiple neighboring intersections according to an embodiment. Fig. 14 is a benchmark of the traffic control system according to the present invention compared with a system using machine-learning, such as neural networks.

In the following, identical reference signs refer to identical or at least functionally equivalent features. DETAILED DESCRIPTION OF THE EMBODIMENTS

In the following description, reference is made to the accompanying figures, which form part of the disclosure, and which show, by way of illustration, specific aspects of embodiments of the invention or specific aspects in which embodiments of the present invention may be used. It is understood that embodiments of the invention may be used in other aspects and comprise structural or logical changes not depicted in the figures. The following detailed description, therefore, is not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims.

For instance, it is understood that a disclosure in connection with a described method may also hold true for a corresponding device or system configured to perform the method and vice versa. For example, if one or a plurality of specific method steps are described, a corresponding device may include one or a plurality of units, e.g., functional units, to perform the described one or plurality of method steps (e.g., one unit performing the one or plurality of steps, or a plurality of units each performing one or more of the plurality of steps), even if such one or more units are not explicitly described or illustrated in the figures. On the other hand, for example, if a specific apparatus is described based on one or a plurality of units, e.g., functional units, a corresponding method may include one step to perform the functionality of the one or plurality of units (e.g., one step performing the functionality of the one or plurality of units, or a plurality of steps each performing the functionality of one or more of the plurality of units), even if such one or plurality of steps are not explicitly described or illustrated in the figures. Further, it is understood that the features of the various exemplary embodiments and/or aspects described herein may be combined with each other, unless specifically noted otherwise.

The present disclosure introduces a control unit that can be applied to traffic signals such as traffic lights in order to maximize the traffic flow. It is noted that maximization of the traffic flow is only an example of a cost optimization, which may be actually performed. For example, the optimization may include minimization of certain traffic parameters such as time of travel, which substantially corresponds to maximization of the traffic flow. Other cost functions may be applied and the optimization (minimization or maximization) of the function may be performed.

The control unit may work on any traffic metric (cost function), which it collects in real-time (i.e. in the form of a data stream), and based on which it outputs a time configuration to be used by the next traffic light cycle (e.g., the sequence of times for red/green/yellow for all traffic lights in an intersection). As becomes apparent in the forthcoming discussion, the unique and novel characteristics of the approach of the disclosure are: the capability to adapt without any cost of pre-training to any intersection, the capability to model any constrains for traffic optimization search, the capability to model the traffic based on any metric in real time, and the capability to offer online traffic control exploiting the spatial and temporal characteristics of the traffic-flow.

In most cases, the real-time traffic prediction and control is based on an optimization engine. This allows the system to adapt to any intersection automatically, creating a set search states, each having traffic predictors for each traffic light and for each timing search domain considered. Each such local predictors is updated in real-time based on specific policies (e.g., time domain proportionate based on the streaming data, update based on selected time slots pre-set by the deployment). Finally, based on the predictors’ output, a combination of traffic signal/lights timings is selected given configured constraints (e.g., max cycle time for the traffic light).

The problem of traffic optimization has been a challenge for many years, with an increased impact every year due to the increase in cars on the roads and congestions on the streets that incur major economic losses and high pollutions in the cities. Many options were studied so far on how traffic can be modelled such that the traffic congestion can be optimized based on the traffic patterns. Most of the employed models and solutions proposed fail despite their complexity and sophistication, because traffic is typically very agile and the behavior changes rapidly, particularly when there are events (e.g., vehicle-vehicle collisions or vehicle-pedestrian collisions, non-allowed second row vehicle stop or parking or the like) on the road. Moreover, each intersection in a city will typically have a different layout and/or flow, which may limit greatly the deployment of specific solutions. Because of these, most solutions to optimize traffic and control the traffic lights times to reduce traffic jams or increase the flow did not work properly.

The present disclosure addresses this issue by tackling the key problems that lead to past failures: rather than trying to learn past patterns, the approach aims to forecast the traffic based on the current context, continuously adapting to the latest evolution, and based on this search in a space for potential traffic times for the one that maximizes the forecast. The control and the feedback about how the traffic actually changed given the past traffic configuration choice, and the impact for the future search is encapsulated in the system that works online, deployed on the edge (e.g., traffic controller).

Concretely, in order to ensure that traffic flow is predicted accurately and new traffic light timings are emitted correctly and with low latencies, a stream operator for predicting traffic data is required that is able to handle large scale scenarios (e.g., towns like Beijing which, in 201 1 , had up to 43 intersections per km squared www.beiiinqcitylab.com). even when operating at high frequencies. To meet these criteria, a dedicated mechanism is developed which performs computationally and resource efficient the following tasks: 1 ) generate search states adaptively based on user configuration and intersection layout; 2) model the incoming metrics based on the search states; 3) generate estimation and prediction of traffic flow time series exploiting both spatial and temporal data; 4) search for parameter and predictor model update in real time; and, finally, 5) generate an optimal traffic light timing sequence. This requirements are fulfilled by the present disclosure as detailed in the forthcoming discussion.

Further, the flow or the traffic flow at an intersection may be impacted also by the layout of the neighboring intersections, referred to as topology, topology map, or spatial topology map. Moreover, the traffic flow of the lanes of an intersection may be affected further by the day, time, and duration of the occurring traffic. In other words, the traffic flow at an intersection may be correlated spatially and/or temporally with the traffic of other neighboring intersections.

One way of controlling the traffic of vehicles at an intersection is by means of traffic signals, which control for example each of the traffic lights of the intersection. This means that not only the type of the light is controlled (e.g., green and red light, eventually yellow/orange as well), but also the respective time (i.e. duration) of each of the red or green light phase.

In many traffic control systems, the traffic lights are controlled based on a fixed time allocation. This means that over the course of one traffic light cycle, each traffic light is assigned a fixed predetermined time period (i.e. duration). This is illustrated by example of Fig. 1 , where an intersection includes four traffic lights, each with three lights red, green, and yellow. The lights may be active along or combined (i.e. illuminated) as controlled via an corresponding traffic signal. In particular, for a traffic light cycle Red -> Red-Yellow -> Green -> Yellow (and cyclic periodic) shown in Fig. 1 , each phase of the traffic lights has a fixed equal time, i.e. the same time duration where the light (or light combination) is active. This predefined traffic light timing is irrespective of the traffic volume (e.g., the number of vehicles) at the intersections or in each lane of the intersection, respectively. This method of fixed traffic light timings may hence lead to inefficient operations by having a low throughput of passing cars and spending most of the times in traffic jams. The major disadvantage is that it does not account for how the traffic evolves at one intersection 1 10 and neither at other intersections in neighborhood of intersection 1 10. In other words, dynamical changes of the traffic in the environment are not accounted for in order to control the traffic at intersection 1 10.

Figure 2 shows an exemplary embodiment of the present disclosure where a certain intersection is equipped with various sensors (e.g., cameras, count control, speed control, etc.) based on which the traffic light sequence may be adapted. In other words, the timing of each traffic light may be altered depending on the dynamically changing environment of the intersection. In the example of Fig. 2, the timing may be 5s red, 3s red-yellow, 7s green, and 4s yellow (and cyclic).

Different approaches are used to select the traffic lights based on certain criteria, for example, the model, rules, and/or policy. Despite the complexity inherent to many traffic flow models, they rely mostly on the presumption that the traffic flow, or respectively the data representing the traffic may be described in terms of a time series. The term time series refers to a sequence of data points that may be indexed in a time order. The data sequence i.e. the data points may be determined at time points which may be temporally equal or non-equal spaced. The data (here traffic data) may be measured or calculated data. In general, the measured/calculated data are continuous in time and become a time series by considering only the data at certain discrete time points. For example, a continuous traffic data may be those measured by one or multiple sensors positioned at a highway or intersection (e.g., a camera) providing data on the number of cars, instantaneous speed or the like. In other words, (traffic) data are generated continuously in a stream-like manner. This makes time series a suitable tool to model traffic flow data, i.e. the stream of traffic data and its temporal evolution in general.

The On-The-Fly data processing of the incoming streaming data is performed via stream processing engines. These engines are an alternative to the approaches that were developed so far for traffic optimization that aim to pre-model the traffic off-line. The traffic data is typically referred to as events and represents a pairing of different pieces of data, which may have a different logical meaning. For example, the data may be an n-tuple representing a lane occupancy on four adjacent roads in an intersection. Such data is/was generated and received in the system in a certain time order. The logic of the processing which may include prediction and control is typically handled by a specific triggering function. In case of traffic, the triggering function may be the arrival of a new event, for example the current number of cars on the lane.

The problem with the real-time processing of continuous streaming data (not necessarily traffic flow data only) is infinite. As a result, an attempt to store the data, as in case of traditional methods, is intractable in terms of the required resources (e.g., memory storage) or in terms of response time, i.e. the time it takes to compute a result to predict the traffic flow. Reactiveness. Moreover, a traffic-flow model for representing streaming traffic data accounts for both the spatial and temporal evolution, which may occur on multiple length scales in space and time. In other words, traffic flow modeling for prediction and control of traffic affects inherently multiple scales. Finally, since the data stream evolves in time, a traffic model (e.g., a learning model) should adapt to different changes in the stream (i.e. concept drift) in order to maintain high prediction accuracy. This will also impact the quality of the control output, for example in terms of how reliable the prediction is. A model not capable to adapt to latest traffic situations, may make the prediction of the flow model obsolete even in a matter of minutes. Further, as data quantity increases in time, the required resource budget increases as well up to a point where it can make the computation too costly. Hence, requirements related to real-time processing may become even infeasible, because traffic prediction and control may not be performed faster than a time scale characteristic for changes of the traffic data. Hence, for traffic control applications where different data arrive in real-time, these data must be processed efficiently to determine, for example, a time configuration in terms of timing signals in order to control the next traffic light cycle.

From the perspective of stream engines, a major problem is that existing technologies do not provide generic solutions for implementing computations and adaptation procedures, typically employed in spatial-temporal prediction and control, with very low latency over event windows. This would imply to enable the (sometimes simultaneous) computation of a sequence of functions (e.g., estimation, parameter search, model update, sequence generation) over large observation time windows, while still preserving the timing and resource constraints. Additionally, they do not offer any support to model the incoming events based on multiple dimensions (e.g., that could be mapped to search states for different traffic lights and timings).

Figure 3 illustrates one of the problems that the present disclosure solves. The time series prediction and control (i.e. estimation, parameter search, model update, sequence generation) must be recomputed over windows of events that slide as new traffic metrics are read (i.e. progress with the stream and might share events between successive instances within a given horizon as only recent traffic metrics are typically relevant due to the highly dynamic behavior of the traffic). Relevant for this problem is spatial-temporal modelling and input-output mapping (i.e. convert traffic flow estimates to traffic light timings). As detailed below, the approach of the present disclosure uses an efficient spatial-temporal model for the prediction and a fast sequence generation for output, which avoid performance degradation as the resource and computation costs grow linearly with the number of elements to be aggregated, as illustrated at a time T4.

In time series modeling of streaming data, data at a later time is predicted. As an implementation option for the predictors that model the traffic for each search space (for each traffic light and for each discrete time domain considered), one can consider modelling the stream events within some given boundaries (e.g., 2 hours of observations on the number of cars preceding the current time). The content of these time windows varies in time as new events arrive and old events fall out of the boundaries of the window and are removed. This means that within such an observation window data used for the prediction are permanently updated. These updates need to be reflected in the function results instantly in order to guarantee correctness. This implies that an accurate prediction of a time series traffic metric, for example, the number of cars is crucial for the subsequent calculation of one or more control signals, such as the traffic light timings.

Typical statistical and machine learning models for time series prediction, including the Auto Regressive Moving Average family (i.e. AR, MA, ARMA, or ARIMA), Bayesian Inference, Regression Trees, and Neural Networks can only model and predict a single dimension time series. In such predictor models it is basically assumed that the correlation in the data can be adequately captured by parameters, which are globally fixed temporally. Furthermore, they are not intrinsically extensible to multivariate predictions making them inadequate for those cases in which the correlations among data are dynamic (time) and heterogeneous (space). This is prevalent in road traffic data.

Various approaches for traffic modelling and control, such as Macro-/Microscopic Models, Filtering Models, Neural Networks, and other combination models have been developed since the 1980s. Figure 3 illustrates the case where a neural network NN is used for traffic flow prediction at an intersection 410 involving four roads S1 to S4, each with multiple lanes and traffic lights. At a certain time, the road traffic is such that with respect to the roads S1 to S4 two cars are on S 1 , six cars are on S2, four cars are on S3, and three cars are on S4. This 4-tupel [2, 6, 4, 3] representing the traffic in terms of a traffic metric“number of cars on road Si” is provided as input data to a neural network 420, which provides a sequence of traffic light timings for S1 to S4. Specifically, in the example of Fig. 3, the timings of the traffic lights corresponding to the GREEN phase (i.e. permitted driving of the vehicles) are 10s, 30s, 25s, and 15s, respectively.

Neural networks as sub-class of machine learning ML models need to be trained first before any prediction can be made. The training itself relies on past data (here past traffic data), which usually have a rather large volume in order to reflect many traffic situations. The NN is then trained in terms of adjusting multiple internal weights of the connections between the multiple internal network layers (i.e. hidden layers) such that a best output is provided at the output layer. The term“best” means that some cost and/or penalty function is optimized. Hence, once the network is trained, they utilize“past experience” to make a prediction when a new state of the world is inputted to the network. In Fig. 4, such data input is provided to the input layer of the network 420.

As evident from this functional paradigm of neural networks or other approaches that utilizes such kind of learning, once the training is finalized the mapping of the input data to output data is fixed, because the weights of the network remain unchanged after the training phase. This is often referred as offline traffic models. The term“offline” means here that the traffic model itself (i.e. the model parameters) is adjusted based on past traffic data. Hence, the respective model with respect to its parameters is a static model, wherein the parameters are not adaptive to the spatial-temporal behavior inherent to streaming traffic data. Another general problem is that NNs or ML models need to be trained for each intersection individually using traffic data reflecting possibly all the traffic situations characteristics for the particular intersection.

Alternatively, one can aim to use stream engines to monitor the traffic. Such approaches can work for some types of modelling functions, but require the re-computation over the window state (corresponding to a time window of a predetermined time duration) for maintaining a snapshot of the traffic flow observations for each incoming event. This obviously affects the real-time constraints and resource usage when scaling to high-frequency streams (e.g., the rush hour traffic situation in Beijing) and long/large time windows (e.g., more than 20 intersections per km squared).

This research continues to be in the theoretical stage and often suffers from a real-world deployment for various reasons. These are for example due to a low forecast precision, strict demand for data, lack of expressivity, understanding of the internal working mechanisms of models (e.g., internals of neural nets etc.), and/or intolerable time cost for the prediction. The key reason that all these models fail in practice is that they do not fully utilize the unique information of transportation networks, namely the dynamic factors such as the spatial- topological structure of the traffic network (i.e. multiple intersections“network knots” connected via multiple roads“network edges”) and the intrinsic temporal flow of vehicles, along with their correlations.

There is no mechanism, stream operator, or solution that enables a traffic prediction and traffic control which combines simultaneously: 1 ) fixed resource usage through an efficient modelling and parameter search, 2) exploiting the spatially and temporally correlated observations for prediction and control, 3) operating with very low latencies, and 4) adapting to any intersection layout.

These problems are solved by the present disclosure. As a special implementation optimization in order to achieve fast and precise flow predictions and control outputs for traffic signals/lights, the disclosure introduces a novel approach for online learning and update, based on spatial and temporal correlations among connected roads. This enables the system to interconnect multiple such deployed systems and apply updates, considering the real-world spatial dimensionality between the locations of deployment of the systems. Traffic flow prediction and control requires access to such spatial and temporal information for judicious decision making. The approach of the disclosure exploits such information through three main contributions.

First, it employs a new mechanism to enable online prediction by modelling incrementally time series corresponding to traffic metrics (e.g., number of cars per lane). This mechanism is extended to be applied also to time series that evolve across multiple dimensions, such as space, having thus a stream compute and control unit capable to model multi-dimensional time series and provide online predictions. Second, the approach enables model updates, not only based on the sensory observations of the time series to be modelled (i.e. incoming stream of traffic metrics on a road), but also based on spatially neighboring time series (i.e. same metrics from neighboring roads) using a weighting scheme. Therefore, the disclosure provides a stream mechanism able to support multisensory fusion based on a configurable spatial neighboring and to share predictions and local traffic models of the stream unit with other deployments. Third, the disclosure provides a stream mechanism able to maximize the output based on the selected traffic metrics. The system is able to self-configure based on the exploration profile of the search space (i.e. duration of the traffic light cycle), the layout of the process to be controlled and optimized (i.e. number of roads in a cross to control).

In this context, the underlying computations are optimized to be constructed incrementally, updating pre-computed states (i.e. stateful processing). The efficient resource usage and the incremental update enable the solution to predict and control, while updating the predictors (with space and time information) at the same time. Additionally, the approach restricts the cached data to the events that are potentially involved in the incremental updates, thus keeping the memory usage constant. Consequently, the disclosure is capable to provide judicious control of traffic lights (or more generally traffic signals) based on observed traffic metrics from the incoming stream or sub-domains of the stream with sub-second latencies. The present disclosure thus provides a solution to a complex problem, namely a low-latency, resource and computation efficient traffic prediction and control in real-time, without deployment costs to adapt to new layouts (e.g., city maps of a larger city etc).

Being an intrinsically complex process, road traffic flow can be modelled for predictions using spatial-temporal models. In these models traffic data are assumed in the form of spatially distributed time series describing local variations of a global phenomenon. Typically, a window and the processing function to be applied (i.e., in the scenario of this invention - traffic prediction and control) are assigned to be executed on one machine (i.e. edge control device of the traffic lights of an intersection) used to run the stream processing engine. A typical default stream implementation based on window operator would hold all the events {ev1 ,ev2, ..., evN} in a memory and at each triggering moment all elements are (re-)-processed to compute the window functions, as shown in Fig. 5. Here, the measured traffic flow refers to the number of cars passing per lane S1 to S4, and the system predicts the future flow and the control timing for the traffic lights to maximize flow (optimization target). In general, an event can be any acquired piece of data. For example, an event can be a picture or a collection of pictures at a given moment or time interval, capturing the intersection or a part of it (one or more lanes or roads) or the like. However, the event may be also or alternatively some processed information such as a number of cars for each lane or an array of such numbers (amounts) of cars per lane or per street for a certain time interval. For example, for a particular state, the event may be a number of cars, which passed from each of the lanes over the crossing (intersection).

The computations, such as those found in spatial-temporal prediction and control, for large observation time windows can require both keeping a large number of states in the memory as well as the re-computation over a large event windows. This makes it a challenge to keep up with the (near) real-time requirements. As mentioned before, this is a major issue for state-of- the-art solutions as well as for more naive stream implementations of traffic metric modelling. This prevents state-of-the-art solutions to obtain an adequate solution for enabling precise traffic flow prediction and control via statistical and machine learning in the case of stream processing.

According to an embodiment of the present disclosure, an apparatus is provided to control traffic signals at an intersection. The intersection may include multiple lanes with each lane having a traffic signal. The traffic signal for a lane has a duration in time, referred to as traffic signal timing. In an exemplary embodiment, a lane may be further equipped with a traffic light.

The traffic signals may be used, for example, to control the traffic lights by which the traffic flow at the intersection may be steered in an optimized manner.

Traffic flow at an intersection is controlled by acquiring for a state among a plurality of states traffic data. A state is a combination of one or more lanes belonging to the first intersection and a same traffic signal timing of the traffic signal for the one or more lanes.

This means that multiple lanes may be specified by state, i.e. one state variable of the traffic model. In this way, the amount of data that need to be provided by the system to perform traffic flow prediction in real-time is reduced. This is referred to as state reduction. This allows to accelerate the state-based data processing by used of a reduced state space.

As evident from the state definition, a state is also specified in terms of a timing related to a timing of a traffic signal, with the timing being the same for the one or more lanes being part of the state. This means that the traffic signal timing, for example, a green phase of each of the traffic lights of the one or more lanes has the same time (i.e. time period), during which vehicles on the respective lanes are permitted to move over the same time duration.

The acquired traffic data are then used to predict traffic data for each state of the multiple states, I.e. the plurality of states.

Based on the predicted traffic data, a first timing for a traffic signal is determined for each of the plurality of states. Said first timing is determined (e.g., by selection) out of a plurality of predetermined timings of traffic signals by optimizing a predetermined function of the predicated traffic data.

With the determined timings for the states, a traffic signal (e.g., a first traffic signal) may be controlled for each of the multiple lanes according to the determined first traffic signal timing.

In other words, the timings of the traffic signals for the states is determined such that the timing- based control of the traffic at an intersection may be optimized, because the traffic signals steer the vehicle flow on each of the multiple lanes over the course of the timing determined for each state.

Figure 6 shows an exemplary embodiment of the present disclosure for controlling traffic signals, including a traffic predictor 610, a traffic optimizer 620, and a predictor update 630. The traffic predictor 610 is provided with traffic data and a set of states {S}⁽ⁿ⁾ at the input. The provision of input data may occur for example at time point t_n with the index“(n)” referring to a discrete point in time. The said time point t_n may be a predetermined time point or an arbitrary point.

According to an embodiment of the disclosure, the term traffic data is based on a traffic metric. Therein, traffic metric corresponds to a quantity suitable to characterize the traffic flow based on one or multiple quantities that may be measured or calculated using a traffic flow model. Suitable quantities to be used as a traffic metric are, for example, the number of vehicles on a lane, the average speed of one or multiple vehicles on a lane, vehicle occupation of the lane, vehicle queue length of the lane, and/or vehicle average waiting times.

The traffic metric is in general relative to the overall measurements during a complete traffic signal/light cycle interval. For example, in case of the metric“number of vehicles on a lane”, the number means the number of cars passing through an intersection (e.g., a first intersection) per each lane over the course of a complete traffic signal cycle. The term traffic signal cycle are also referred to as signal cycle or simply cycle and are used interchangeably. A traffic cycle refers to a typical time during which a sequence of traffic signal, respectively, traffic light signals for the multiple traffic lights belonging to an intersection (e.g., a first intersection) is swept through completely with their respective traffic/signal light timing. In other words, the complete traffic cycle interval denotes a cycle in which green light is applied in sequence to all states once. For example, in Figure 9, the complete traffic cycle interval would correspond to the cycle of the states s1 , s2, s3 meaning offering green light for the corresponding lanes one after the other, in each of the states. After each state got it turn, the cycle is completed.

Another example of such a metric relative to said cycle, is an average speed of vehicles on a lane passing through the intersection during the traffic light cycle interval. Hence, the quantitative measurement of traffic data according to a metric is a result of aggregating of such data, for example, by counting, summing, and/or other types of statistical averaging. This is performed for each of the states managed by the traffic signal/light system. As an example, assume that a traffic light is assigned 20 seconds of green time for the managed lanes, then the traffic metric for that traffic light is measured for any of the metrics for 20 seconds for the considered lanes.

The green time assigned to one or more traffic lights may also be referred to a phase, a green phase or the like. This means that during a green phase respectively green time it is understood that vehicles on the one or more lanes with their said traffic lights being green are allowed to drive or move on that lane. Hence, the motion of one or more vehicles generate in then end a traffic flow, and hence a streaming traffic data associated with the flow.

In general, the phases related to traffic signals may include Red, Green, and/or Yellow/Orange. The phases Red and/or Yellow/Orange may be used in addition in the traffic flow model for controlling the traffic signals respectively the traffic light timings. The timings for each of the phases may be same or different. In general, the timing of a phase may be different for different traffic cycles. In other words, for any of a cycle the timings may be adapted based on traffic data.

These are different types of possible traffic metrics. In addition or alternatively, traffic metric may be a particular type with the metric referring to the amount and/or magnitude of the respective metric. For example, in case of the particular metric type“number of cars per lane” the metric is provided by the number of cars. Metric may also refer to the number of cars of all lanes belonging to the intersection. The metric may alternatively or in addition be a change of the amount and/or magnitude. This means that the traffic metric is measured or calculated at two or more different points in time in order to determine an actual change of the traffic metric.

As mentioned before, for modeling of traffic-related problems including vehicles on lanes of one or multiple, connected intersections, the flow i.e. the streaming of traffic data is suitably described in terms of time series. This means that the streaming model is used to predict a traffic flow/metric at a future time. A future time may include one time point t_n+i being later than a previous time point t_n or may include N multiple time points {t_n+i , ..., t_n+N} with N > 1 later than t_n.

In an exemplary embodiment of the disclosure, a state belonging to the plurality of states {S}⁽ⁿ⁾ (i.e. the set of states) entails one ore more lanes of the intersection, for example, as a lane label or index“Li”, with i = 1 ,...N referring to the index i of a lane and N being the total number of lanes of the intersection. In this way, a state may specifies the one or more lanes, which are merged into a single state. This is referred to as state reduction. This enables to reduce the amount of data that need to be hold available for optimizing the traffic metric. As a result of the state reduction, the cost of infrastructure may be reduced since less storage is needed. Moreover, using a reduced state space for the traffic flow control enables a faster processing of the data, including traffic data acquisition and subsequent optimization.

The state entails further a traffic signal timing of the lane. The timing may belong to a plurality of signal timings, for example, of {10s, 15s, 20s, 25s, 30s}. As detailed further below, the plurality of timings define part of the search space used in the optimization of the traffic flow. The values of the timings may be equal- or non-equal-spaced. In the above example, the timings are predetermined. Alternatively, the timings may by be adapted with reference to the time point t_n of the data acquisition and/or in accordance with the traffic data.

The term“timing” refers to a duration of the time of the signal. In the exemplary case where the traffic signal controls a traffic light, the timing corresponds to the time period during which one or more of the traffic lights e.g., {Red, Green, Yellow/Orange} are active, i.e. turned on. Alternatively, the timing duration or timing length of a traffic signal may be represented indirectly, for example, by a particular modulation and/or coding scheme. For example, the timing signal length may be realized via a pulse-width-modulation scheme (PWM) or the like.

For a state S ⁿ⁾ out of the multiple states {S}⁽ⁿ⁾, traffic data is acquired. The data acquisition may include traffic data at the latest and/or current time point t_n.

According to an embodiment of the present disclosure, the traffic data acquisition may include in addition or optionally“old” traffic data, referring to traffic data acquired at one or multiple points in time t earlier than the present time point t_n (i.e. t < t_n) In other words, the acquisition of traffic data may be performed using a traffic data history.

With the acquired traffic data and states {S}⁽ⁿ⁾ as input, the traffic predictor 610 predicts for each state traffic data at a later time point t_n+i > t_n, employing at the least the latest traffic data. This refers also to online prediction of the traffic flow. Here, the prediction is performed by calculating the data through a time series model, based on latest and/or past traffic data. This is done for each state among the multiple states. Recall that a state entails also a signal timing among multiple signal timings.

Alternatively or in addition, the traffic data acquisition may be for one or more states out of the plurality of states.

The traffic optimizer 620 then determines for each of the plurality of states a traffic signal timing by optimizing the predicted traffic data. This means that a traffic metric, representing a quantity for the traffic of all the lanes of an intersection, is optimized using a predefined function of the metric. For example, if the metric is“number of crossing cars per lane” (with reference to a cycle), then the optimization refers to maximizing this number. If on the other hand, the metric is “waiting time of a car on a lane”, then the optimization refers to minimizing the waiting time. Depending on the metric or respectively the traffic data, the traffic optimization determines a timing out of predefined multiple signal timings. The predefined function may be, for instance, a function that depends on the metric. For example, If the number of cars passing the intersection per time and/or via a lane is acquired, then the function may be the sum of the cars. The function corresponds to the cost function optimized (here maximized), as described above. For example, the sum of the number of card passing the intersection would be maximized. Another function may be an average or other statistic or any function of the acquired data.

Hence, the prediction and optimization of the traffic flow utilizes a (reduced) search space, which is spanned by the plurality of states {S}, with each state being specified in terms of one or more lanes [L¹, L², ...] and a same timing T_j for a state j. Formally, such as state may be formally written as S_j = {[L_j ¹ , L_j ², ...], T_j}. Note that in case of more than one lane, multiple different lanes carry the same subscript“j” to indicate their association to the common state j. Moreover, as mentioned before, lanes belonging to the same state have the same timing Tj of their respective traffic signals. A state defined in this manner, in which multiple lanes are reduced into a single state reduces the data to be managed by the system, resulting in a speed up of the processing of the multi-dimensional space-time traffic flow data.

The prediction of traffic data is done for each of the multiple states for the respective timing T_j. Assuming for example that there are N lanes and M timings, the search space spanned by the states has therefore a dimension of N x M. The entries of the search space are the traffic data predicted using a streaming model for the traffic predictor.

According to the present disclosure, each state S_j has its own predictor model. For example, if the timings are {10, 15, 20} seconds, then for each timing for which traffic data is predicted a separate predictor model is used.

Alternatively, a group of states among the plurality of states {S} may use the same predictor.

Once a traffic signal timing is determined for each state of a lane, the timing is used to control a traffic signal for the lane. According to the present disclosure, the control of the signal according to the determined timing means that the control signal has a time length corresponding to a time width of the signal linear proportional to the timing. For example, assuming that the signal width is 2s for a 2s timing, the signal width is 4s for a 4s timing.

Alternatively, the signal width may be a linear ratio of the timing, wherein the signal width is scaled with the timing. These types of mappings between the timing and the signal may be referred to as pulse-width modulation PWM. Alternatively, the signal width may be related to the timing according to a monotonous function, which may not necessarily linear to perform a mapping from the signal timing to the signal itself. Alternatively, the determined timing may be mapped to the signal height. This corresponds to a modulation of the amplitude of the traffic signal corresponding to the timing.

In an exemplary embodiment of the present disclosure, the traffic signal is a signal corresponding to the traffic light. Respectively, the determined signal timing is a traffic signal light timing. In other words, traffic lights are used in order to control the flow of the traffic on a lane by visually signaling through lights (e.g., Red, Green, Yellow/Orange and combinations thereof) when a manned vehicle is permitted to drive (signal for Green light based on timing of Green phase) or should stop (signal for Red light based on timing for Red phase).

In addition or alternatively, the vehicle may receive the signal for driving or stopping according to the signal timing of the lane, for example, via road-vehicle communication using a wireless connection. Based on the received signal, the driver may be informed for example via an acoustic signal (e.g., acoustic speech instruction or different types of acoustic signals to indicate a Green or Red phase). The driver instruction for performing control of the vehicle according to a Green or Red phase may be also indicated through mobile phones, smartphones or the like, which perform wireless communication with the intersection. In general, any form of road- vehicle communication to transmit the traffic signal according to the traffic signal timing may be used. Hence, even the traffic flow involving mere autonomous vehicles (manned or unmanned) or a combination of autonomous and non-autonomous vehicles may be performed.

According to an embodiment of the disclosure, traffic data is optimized according to a constraint including a maximum cycle time of the traffic signals. However, the present invention is not limited to such constraint. Rather, other alternative or additional constraints may be used such as minimum green time per lane (corresponding to the minimum time period for green light configurable), yellow time (corresponding to the time period in which the yellow light is active), or constraint that there is always a green light for turning right (there are intersections where turning right is always permitted. In those cases sometimes there is a traffic sign or a one light traffic with the color green always lighten up).

As an example of a time constraint, when 8 bits are used to encode the timings, then the total duration of a complete cycle is typically limited to 255 seconds, with 2 digits used only for the countdown for the green time. As a result, each direction of a lane cannot get more than 99 seconds green.

The term traffic light cycle refers to the complete execution of a traffic plan, where all the states defined for each of the direction/groups of lanes are executed. For example, imagine a 4 cross intersection. A possible plan is to give a green time for each of the directions in a clock-wise order, starting with the direction north, east, south, and then west. This is a complete cycle or simply cycle. Of course, the phases of the traffic light plan may be more complex, and beyond the number of directions in an intersection.

By having a maximum cycle time as part of the optimization, the timings allowed for the multiple timings defining the search space is restricted. In other words, a signal timing may not exceed the maximum cycle time. As a result, the size of the spanned search space is reduced. This enables a fast search for a signal timing for each lane and hence a fast optimization of the traffic data using a predetermined function.

According to an embodiment of the disclosure, the constraint is predefined or selected with reference to the traffic metric. This means that, when optimizing the traffic data, a predetermined constraint may be required for performing any optimization irrespective of the used metric (hard constraint). In turn, selecting a constraint depending on the traffic metric allows for adaptation of the optimization (soft constraint, i.e. selected). Since the traffic data optimization is performed using a predetermined function, the possibility to select a constraint that may be more suited for the to be optimized function provides the flexibility to tune further the optimization performance of the traffic optimizer 520 (e.g., in terms of convergence speed to find a functional optimum etc). For example, it may be decided on a contraint such that one does not want to allocate more than 99 seconds green for a direction (predefined limit) or one does not want directions to have a difference more than 20% on average among each other, with the times being defined based on the traffic metric collected).

In the exemplary embodiment of the present disclosure shown in Fig. 6, a traffic constraint may be provided to the traffic optimizer 620. In addition or alternatively, traffic optimizer 620 may be provided with more than one constraint. These constraints may be predefined and/or selected based on the same traffic metric or different metric. In other words, the optimization of the traffic data may be a multi-constraint optimization. This allows the determination of a signal timing for a lane that simultaneously meets requirements for an optimal traffic flow based on multiple constraints.

According to an embodiment of the present disclosure, the multiple states (i.e. the state set {S}⁽ⁿ⁾) include states belonging to lanes of a second intersection, which is spatially neighboring the first intersection. This means that the first and second intersection have a certain spatial distance from each other. This distance may be determined, for example, based on a spatial map including both intersections.

According to an embodiment of the present disclosure, the acquisition of traffic data for a state includes traffic data of a state, belonging to a lane among multiple lanes of the second intersection. Since the first and second intersection are spatially apart, the first and second intersection are different.

As a result, the prediction and subsequent optimization of traffic data for the state of a lane accounts for spatial correlations among the traffic data streams of states belonging to lanes of different intersections. This is because the prediction of traffic data for a state uses the acquired traffic data, which may in general include data from all states, i.e. states of all intersections.

As mentioned before, the data acquisition includes actual traffic data, but also past traffic data based on a traffic data history. Therefore, the approach of the present disclosure accounts also for temporal correlations among the traffic data streams of states belonging to lanes of different intersections.

Consequently, the prediction and optimization of traffic data for a state of a lane accounts simultaneously for both spatial and temporal correlations of traffic data streams among all the lane states of all intersections in general. These spatial-temporal correlations include short- and/or long scales in space and/or time, depending on the distant range of states belonging to neighboring intersections and/or the temporal length of the traffic data history. Hence, depending on the strength of spatial-temporal correlations (i.e. the signed and/or absolute value thereof), cooperative effects inherent to traffic data streams are accounted for and ultimately impact the future evolution of traffic data for a single lane, i.e. the predicted and optimized data.

According to an embodiment of the present disclosure, the prediction of traffic data is based on a parametric prediction model. A parameter of the prediction model may be updated in accordance to the acquired traffic data. The parameter may be updated also according to predicted traffic data and/or in accordance with the spatial neighboring of the intersection.

The use of predicted traffic data allows an update of a parameter of the predictor model in case of missing traffic data. For exmaple, it may happen that the data acquisition for a state terminates and provides e.g., corrupted data, as may happen in case of a power failure of a camera, which is presumed to provide data for one or more lanes based on a traffic metric. In this case, the predicted traffic data is considered as real-data for the parameter/model update. Furthermore, the official count will be available only for the previously selected model.

Hence, the model used to predict traffic data for a lane is a dynamic prediction model, whose parameters may be adapted (adaptive predictor model). In other words, the predictor itself is adaptive and an integral part of the approach of the present disclosure. The parameter update of the model in turn is based on traffic data related to the intersection (i.e. first intersection) and/or using other traffic data from lanes of other distant intersections (i.e. one or more spatially neighboring intersections).

In an exemplary embodiment of the present disclosure, the update of the predictor model may include more than one parameter, and may depend on the complexity of the specific model used to perform the prediction.

This update or refresh of one or more parameters of the predictor ensures that the parameter representation of the predictor continues to be up-to-date with the current traffic data of the multiple states, i.e. the acquired traffic data.

As mentioned before, the acquired data may include also data acquired earlier, so that the parameter update accounts also for history effects of the traffic data. In addition, a parameter may be updated according to the most up-to-date traffic data, namely the predicted traffic data.

Moreover, the parameter refresh with regard to the spatial neighboring of other intersections allows to include also spatial information of the traffic data into the parameter representation (i.e. the parameters) of the prediction model. As a result, the parameter update of the predictor model itself of the present disclosure enables to account for spatial-temporal (ST) effects (i.e. ST-correlations) also in the model parameters, with the result that the predictor is inherently dynamic through the adaptation of the parameters. This way of online updating of the model allows to predict and to optimize traffic data, and hence to determine a traffic signal timing for a lane with a high accuracy. This is not only because of the data used for the prediction are being up-to-date, but also because the predictor model itself is permanently updated.

This is in striking contrast to predictors using machine learning (ML) models, such as neural networks as shown in Fig. 4, which rely on an offline pre-training using“old” data, i.e.“old” trained parameters of the network are used to predict traffic data. This means that in NN-based approaches to model streaming traffic data the parameters are not kept up-to-date in order to account for dynamic and spatial changes inherent to traffic data. This is provided by the parameter update of the present disclosure.

According to an embodiment of the present disclosure, the parameter of the prediction model of a first state based on a second state selected among the plurality of states. In other words, a parameter update is performed on a state-to-state basis, involving different states. This allows for a fine-tuned parameter adaptation on a state basis, by which the prediction of traffic data is performed with a further enhanced precision.

According to an embodiment of the present disclosure, the selection of the second state is based on a selection policy. The policy may be a state for which traffic data has been acquired, a state for which traffic data has been predicted, and/or a state whose traffic data has been obtained by transforming a traffic metric of a different state.

The above selection may hence be cast into any of a selection policy, for example,“state traffic data acquired”, “state traffic data predicted”, and/or“state transformed data”. These policies may not be limited to those listed above.

For example, in addition or alternatively, a policy may be related to states that were not selected respectively determined from the multiple states when performing the optimization of the traffic flow function. To illustrate this point, suppose that a state e.g., of Lane 1 is determined with a timing of 20 seconds (instead of 30 seconds) as a result of the optimization. This means, after the elapse of 20 seconds for that states, one receives after the cycle a traffic metric corresponding to the state {Lane 1 , 20s}. These traffic values are to be filled with values following a policy. This policy may be: (1 ) no update if the state was not selected, (2) time proportionate values based on the value for the state for which the metric has been received, (3) other functions that transform the read metric for one state to the value corresponding to another state, (4) update with the forecasted value (i.e. a predicted value), and/or (5) a combination of any of these. For example, make an update if no value was received during the last 10 minutes and the update is done with one of the options (2)-(3)-(4).

In the exemplary embodiment of the present disclosure shown in Fig. 6, the updating of the parameters is performed by predictor updater 630. As illustrated, updater 630 receives as input acquired traffic data and/or predicted traffic data (i.e. the output of traffic predictor 610) and/or other traffic data corresponding to data of lanes from neighboring intersections. The updater 630 outputs the updated model parameters and provides these as input to traffic predictor 610. In the predictor 610 the updated parameters are used in the model to predict traffic data for the next cycle.

In the exemplary embodiment of Fig. 6, the predictor 610, optimizer 620, and updater 630 are separate units and/or separate circuitries. Alternatively, units 610, 620, and 630 may be assembled in one common unit and/or part of the same circuitry. The circuitries may be assembled further on the same board or different boards in order to enable a modularity of the system.

According to an embodiment of the present disclosure, a method is provided for traffic signal control at a first intersection including a plurality of lanes and a traffic signal for each of the plurality of lanes, comprising the steps of acquiring traffic data for a state among a plurality of states, wherein the state is a combination of one or more lanes belonging to the first intersection and the traffic signal for the one or more lanes being same over a traffic signal timing out of a plurality of predetermined traffic signal timings; predicting traffic data for each state of the plurality of states based on the acquired traffic data; determining for each state among the plurality of states a first traffic signal timing out of the plurality of predetermined traffic signal timings by optimizing a predetermined function of the predicted traffic data; and controlling a first traffic signal for each of the plurality of lanes according to the determined first traffic signal timing.

According to an embodiment of the present disclosure, a computer-readable non-transitory medium stores a program, including instructions which when executed on a processor cause the processor to perform the steps of the method for traffic signal control.

In an exemplary embodiment of the present disclosure, an architecture of a processing and control system that provides a solution for traffic optimization in real-time according to the present disclosure as shown in Fig. 7. As apparent from the previous discussion, the approach is going beyond the traditional approaches of considering prediction and control sequence generation as a simple stream processing. At its core, the disclosure provides a control unit that can be applied to traffic signals (e.g., traffic lights) in order to maximize the traffic flow.

According to the exemplary embodiment of the present disclosure shown in Fig. 6, the control unit may include the traffic predictor 610, traffic optimizer 620, and the predictor updater 630.

In the initialization/configuration, the disclosure can adapt to the user constraints and the intersection layout (and specific of the traffic signals and/or lights). In other words, the approach allows to account for spatial effects of the traffic data due to the spatial location of and connection among other intersections (i.e. the layout, topology of the street/road network, or city plan). The approach may work on any traffic metric, which is collected in real-time. This traffic metric (i.e. the acquired traffic data based on the metric) is used to generate a time configuration (i.e. the traffic signal/light timings) by means of optimizing the traffic data and the timings are used for the next traffic signal/light cycle.

In view of Figs. 6 and 7, the unique characteristics of the present disclosure may be summarized as follows:

1. The approach is generic as it can work with any available traffic metric (e.g., number of cars, speed of cars, occupation at traffic light, etc).

2. The approach is general as it is applicable to any intersection layout, without any prior knowledge or design, by simply providing the number of states of the traffic signal/light to be controlled (i.e. number to traffic signals/lights for all streets in an intersection).

3. The approach uses online statistical and machine learning techniques to optimize the traffic. The approach does not require any prior training or deployment/configuration efforts (i.e. major benefit to reduce the cost to setup and deploy a traffic solution).

4. At deployment time, all the necessary specific configurations for controlling the traffic signals/lights are created by the control unit. After this initial configuration stage, the control unit, will work continuously in the sequence: (a) collect traffic metric, (b) process it, and (c) output a configuration for the traffic signal/light. The underlying processes of the present disclosure are introduced in the following:

(1 ) Based on the initial setup (i.e. number of traffic light states and time domains considered), a collection of search states will be created, one for each pair of traffic state and traffic domain.

(2) For each search state, a traffic modeller (i.e. online time series predictor) is used.

The approach may accept various prediction algorithms or methods as long as they are able to provide a prediction based on the inputs that they model.

(3) For each traffic metric that is received, the following processing sequence is then repeated. Typically a new traffic metric can be received either after each traffic signal/light cycle, or after several such cycles, or at certain fixed times.

(a) The traffic metric values, are used to make an update in the modellers (i.e. predictors) of the search space. Multiple strategies may be used to select which modeller should be updated (e.g., all of them, the last ones selected).

(b) The modellers from all search spaces will be queried to make a prediction on the traffic metric they model.

(c) Based on the predictions, a combination of times per traffic signal/light is selected from the search collection, in such a way to: (i) optimize (e.g., maximize) the predicted outcome for the traffic; (ii) meet a time constraint set for the traffic signal/light cycle.

(d) The traffic cycle time (i.e. the constraint) may be pre-configured, or dynamically selected based on the traffic metric.

(e) The selection is recorded and the corresponding times for each traffic signal/light is generated and used for example for the traffic light.

This may be appreciated further by considering the most representative technologies and concepts that relate to the present disclosure:

A. Stream Processing

Stream engines (e.g., Flink https://flink.apache.org , Spark Streaming, Storm, Samza, and Dataflow) are the main stream technologies are related to the present disclosure. Stream engines have the role of processing data on-the-fly (in movement). They provide computing capabilities based on the time ordering of the stream. Depending on the specific engine, the time can be further set to refer to event time, processing time, computer time or arrival time of the events. Most of the stream engines allow some form of grouping the events in windows. Depending on the API of the stream engine, different flexibility levels to define and to drive the computation on the window exist.

None of these engines offers dedicated operation support to handle traffic metrics, model the time series across multiple dimensions, or build online traffic optimization support.

As evident from the previous discussion, the present disclosure provides all of these capabilities.

B. Statistical and Machine Learning ML for Streaming Time series Prediction

Formally, a time series is a sequence of historical measurements of an observable variable at equal time intervals. Time series are studied for several purposes such as the forecasting of the future based on knowledge of the past, the understanding of the phenomenon underlying the measures, or simply a succinct description of the salient features of the series. The forecasting domain has been influenced, for a long time, by statistical learning methods. The objective of ML methods is the same as that of statistical ones. They both aim at improving forecasting accuracy by minimizing some loss function, typically the sum of squared errors. Their difference lies in how such a minimization is done with ML methods, utilizing non-linear algorithms to do so while statistical ones linear processes.

ML methods are computationally more demanding than statistical ones and come in different approaches, from Multi-Layer Perceptron (MLP), to Bayesian Neural Networks (BNN), or Radial Basis Functions (RBF), and from CART Regression Trees (CART) to Gaussian Processes (GP) or Long-Short Term Memory (LSTM) recurrent networks. Despite having such a diverse pool of models, the most relevant aspects to consider in evaluating and designing a time series forecasting system are the accuracy, the goodness of fit, and the computational complexity of the method.

For such an analysis, one must consider how the model can:

(a) handle uncertainty (i.e. specify or not a probabilistic model of the data generating process);

(b) specify structure (i.e. assume additivity of predictor effects when specifying a model), and (c) include empirical evidence (i.e. allow or not high-order interactions not pre-specified at design time). Moreover, for the traffic control scenario, a prediction system should fully utilize the topological structure and intrinsic temporal flow of vehicles and their correlations.

In order to situate the present disclosure, some relevant research work is further looked at and it is emphasized where the approach of the disclosure is different from state-of-the-art.

The work by Blandin, Sebastien, et al. "On sequential data assimilation for scalar macroscopic traffic flow models." Physica D: Nonlinear Phenomena 241.17 (2012): 1421-1440, considers the problem of sequential data assimilation for transportation networks using optimal filtering with a scalar macroscopic traffic flow model. Properties of the distribution of the uncertainty on the true state, related to the specific nonlinearity and non-differentiability inherent to macroscopic traffic flow models, are investigated, derived analytically, and analyzed.

The major difference from the present disclosure is that there is no streaming deployment, no adaptability based on layout, no modeling based on discrete times, and no spatial-temporal updates across neighbors models.

As evident from the previous discussion of the present disclosure in conjunction with Fig. 6, the approach of the disclosure uses both the spatial and temporal correlations among the time series of traffic metrics to generate a more accurate prediction and subsequent timing sequence generation for the traffic signal/light control.

As detailed already, the learning model (i.e. the predictor model) of the present disclosure used for the prediction:

(a) exploits the inherent uncertainty in the traffic metric fed to the system.

(b) uses only the available data in the stream window and hence a relatively limited sample size for training/learning.

(c) exploits the spatial and temporal evolution in an interpretable model.

(d) can consider and isolate the effects of a small number of variables (i.e., spatial and temporal distance among events from neighboring streets).

In the following, further details are provided to understand the underlying operation principle of the present disclosure. To simplify the discussion, the terms“traffic signal” and“traffic light” will be used synonymously. Figures 8 to 12 illustrate by example of a single intersection how the traffic light timings are determined. Figure 8 shows an exemplary embodiment of the disclosure, comprising two control units 810 and 820 that operate in concert on the incoming data stream of traffic metric events. From a functional point of view, control units 810 and 820 refer to the reduced states S1 and S4 for which a timings is determined. In this example, each control unit 810 and 820 may further include:

(1 ) a stream compute unit capable to model time series (i.e., stream of observations) for online predictions: model updates based on spatial and temporal information received from neighbour predictors,

(2) a stream control unit that using the configuration of an intersection (here the city map), will create a set of predictors for each traffic light state and for each discrete time dimension to be considered, among which it will search for the best configuration while trying to maximize the outcome, and

(3) a stream compute unit capable to optionally synchronize with similar compute units in order to update the predictors model based on neighbour states.

With reference to Fig. 9, we want to solve the problem of optimizing a traffic light, which is part of an intersection. Here, the intersection includes five streets with a total of six lanes L1 to L6. It is further assumed that a full traffic cycle includes three phases, phase 1 to phase 3. The core idea is that, based on the default traffic light cycle, for each phase 1 to 3 the traffic lights having a Green light at the same time are combined and merged into one state. The other traffic lights are paired accordingly. In this way, one creates a minimal number of states. This reduces the number of states and hence the amount of data to be handled by the system for making traffic predictions in an efficient manner in real time. Here, lanes L1 and L5 are combined into a traffic light state cycle S1 , as shown on the right of Fig. 9. As to the other lanes having a Red light, they may be paired in any way because no traffic flow original from these lanes due to the Red light phase, indicating a non-permission of moving a vehicle on the respective lanes L2, L3, L4, and L6. For the other phases 2 and 3, the state reduction is performed accordingly. This defines the states S1 to S3 of the traffic light state cycle. Typically each traffic light control unit will have a state machine for the operation, which the apparatus of the present disclosure uses.

In order to determine the traffic light timings, the present disclosure uses a learning system capable of modelling the traffic flow, searching the space of parameters of the model, update the model, and generate an output corresponding to the timing sequence of the traffic lights. In this example, the traffic metric for the traffic data is the car throughput.

The sequence of operations for the timing determination is graphically depicted in Figs. 10 and 1 1 and consists of the following steps: 1. Start from a default sequence (e.g., 20 seconds green for each light in turn).

2. Build a discrete search space with each state corresponding to a corresponding time allocation.

3. Estimate what would be the outcome from each traffic light as if it would be assigned green for the corresponding time seconds.

4. Select the combination that maximizes the greedy outcome given a constraint (e.g., maximum amount of time for a cycle).

5. Update the estimator model based on the observation to correct the estimation based on latest observations (e.g., different policies can be applied of which predictors are to be updated and how based on the next metrics received. In the example below the policy is to update only the selected states. Other policies can be to update proportionally all states, update based on specific functions, update after a given time passed since last update...)

Each traffic light has a collection of models (i.e. a predictor), one for each state to search and each to be trained for a dimension (i.e. for a specific time assigned for green). In this example, each traffic light has five models for 10, 15, 20, 25, 30 seconds). Recall that this defines the search space for the states, spanned by the plurality of states and the plurality of predetermined timings.

Once the timing sequence that maximizes the expected outcome is selected, this timing is used for the traffic light. A new observation on the number of cars (i.e. a new acquisition of traffic data) that actually passed will be available in the stream. The predictive models is then updated (i.e. their parameters) based on the new sample of traffic data. If the estimation is too optimistic, then having lower observed values, will decrease future estimates. In turn, this will make more likely the selection of other time combinations. This mechanism is depicted in Fig. 1 1.

With reference to the previously introduced example, the approach of the present disclosure may be extended for any intersection configuration and shape, using the number of traffic lights states the intersection has and optionally a configuration of times used, like the default sequence. This decouples the use of a strong prior as in other methods (e.g., Neural Networks, Deep learning, and Deep reinforcement learning networks) and makes the approach of the disclosure more flexible.

Any type of observations of traffic metrics (e.g., numbers of cars passing, length of queue, waiting time, and average speed or the like) may be used to update the predictors, which is not possible with other approaches. Figure 12 introduces the mapping from the physical setup to the actual instantiation of the invention.

In large scale scenarios, such as a city, each traffic light in an intersection intrinsically influences the streets to which it is connected. In other words, the evolution of the local traffic flow is affected by the traffic farther away. In such a case, the disclosure exploits the topological structure and intrinsic temporal flow of vehicles and their correlations (spatial-temporal correlations). Each intersection has a predictor model and the knowledge of neighbouring streets is used to locally update the model.

Figure 13 shows an exemplary embodiment of the present disclosure, with a large-scale instantiation of the system. Supported by the previous instantiations of the disclosure, further details on the implementation of the disclosure are provided in the following.

The first aspect refers to the online prediction on stream data for traffic control. The initial implementation of the system considered Statistical and Machine Learning algorithms that were developed by porting traditional models (e.g., Auto-Regressive Moving Average family: AR, ARMA, ARIMA) to work in an incremental fashion. The model coefficient calculations implied by such models were recursively estimated using iterative Kalman Filters, to improve computation performance and avoid expensive optimization techniques (e.g., Least Squares). Other predictor models, or even black-box algorithms can may loaded by the system.

The second aspect targets the stream operator used for the prediction. In order to implement the incremental predictor models, the system extends the classical architecture of a stream operator (e.g., process function) to support incremental computation of the online prediction functions, which were implemented in a library of user defined formulas. This allows a flexible use of different predictor models. States were added in order to ensure fault tolerance and robust incremental processing.

The third relevant implementation detail refers to the adaptive stream operator for traffic control. In order to provide the adaptive component to any intersection layout, an own stream operator architecture has been built to accept an input configuration (e.g., number of traffic light states, number of discrete times states) such that in the initialization phase of the operator a set of predictors are be created, for each such combination (e.g., a matrix of predictors of size: number of traffic light states multiplied by number of time states).

Each observation of a traffic metric (i.e. traffic data acquisition) is used to update the model predictors in real-time. The update can be driven according to policies (e.g., update only the predictors corresponding to the times that were selected, update all proportionally based on times, update based on some distribution functions across all predictors). Moreover, at each synchronization of neighbourhood traffic control units, the other models’ predictions are used to update the local copy of the neighbours and the corresponding weights, which will be used when the local prediction models are updated. This accounts for a sensor fusion mechanism.

In the overall architecture, each predictor is triggered to generate an estimation of the traffic metric considered in the prediction modelling, of all combinations of times for each traffic light. From this combination, using dynamic programing, the optimal combination (of the timing) that maximizes the metric is selected.

Finally, in order to support the spatial-temporal models update, the predictors are extended to consider in the parameter update a weighting mechanism that is given by the neighbourhood distance and correlations. The incremental formulas of the prediction models (i.e. of the time series model) are then extended to support also the weighting of the parameters of neighbourhood models. This allows for a spatial-temporal update of each predictor. This enables further the sensor fusion in order to extend to multiple scales and to improve the quality and accuracy of the predictions.

In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus, and method may be implemented in other manners. For example, the described apparatus embodiment is merely exemplary. For example, the unit division is merely logical function division and may be other division in actual implementation. For example, a plurality of units or components may be combined or integrated into another system, or some features may be ignored or not performed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections may be implemented by using some interfaces. The indirect couplings or communication connections between the apparatuses or units may be implemented in electronic, mechanical, or other forms.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one position, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

In addition, the functional units in the embodiments of the invention may be integrated into one processing unit, or each of the units may exist alone physically, or two or more units are integrated into one unit. Embodiments of the invention may further comprise an apparatus, which comprises a processing circuitry configured to perform any of the methods and/or processes described herein.

Embodiments of the predictor 610 and/or optimizer 620 and/or updater 630 as shown in Fig. 6 may be implemented as hardware, firmware, software or any combination thereof. For example, the functionality of the any of the predictor, optimizer and/or updater may be performed by a processing circuitry with or without firmware or software, e.g., a processor, a microcontroller, a digital signal processor (DSP), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or the like.

The functionality of the predictor 610 and/or optimizer 620 and/or updater 630 may be implemented by program instructions stored on a computer readable medium. The program instructions, when executed, cause a processing circuitry, computer, processor or the like, to perform the steps of the traffic signal control method. The computer readable medium can be any medium, including non-transitory storage media, on which the program is stored such as a Blu ray disc, DVD, CD, USB (flash) drive, hard disc, server storage available via a network, etc.

An embodiment of the invention comprises or is a computer program comprising program code for performing any of the methods described herein, when executed on a computer.

An embodiment of the invention comprises or is a non-transitory computer readable medium comprising a program code that, when executed by a processor, causes a computer system to perform any of the methods described herein.

According to an embodiment, a computer-readable non-transitory medium stores a program, including instructions which when executed on a processor cause the processor to perform the steps of the method.

BENCHMARKING

As evident from the previous discussion, the present disclosure introduces a new approach for stream-based prediction and control of traffic signals (in one exemplary implementation these related to traffic lights) and city traffic based on analytical online predictions. It enables online modelling, update and prediction capabilities of traffic time series in multiple dimensions (i.e. space and time). The system includes a stream compute unit, which is able to carry out such processing on real-time streams of traffic metric data, with a millisecond-level latency.

The system enables online updates and fusion of prediction models based on spatial and temporal information of similar predictor models, including a stream compute unit capable to interconnect with other such units, for synchronization of the prediction models. Moreover, the system is capable of self-configuration at the level of the control unit for traffic signals respectively traffic lights, based on the intersection configuration and layout, by creating distinct predictors for traffic signal states and for each time-dimension to be explored. This allows for a generic traffic optimization control unit adaptable to any type of traffic metric.

As discussed before, functionally, the system is capable of online selection of traffic signal times (for control of traffic signals that may active and/or drive traffic lights) based on predictions, while maximizing the selected traffic metric out of the configurable ranges of timing options (i.e. the search space of the states). The architecture of the present disclosure may be deployed on the edge device, on the traffic controller, or in a central control centre (e.g., a cloud).

In order to assess the capabilities of the present disclosure, a use-case scenario is introduced. With this example it is aimed to emphasize the capabilities of the stream operator underlying the invention and its performance on real-world traffic data. As simulation environment the Simulation of Urban Mobility (SUMO, version 0.32.0 - https://su o.dlr.de) is used and is tested on a PC with 24G of RAM.

As mentioned already, in order to pre-evaluate the benefits of the disclosure for business scenarios, such as traffic flow optimization, a standardized simulator is used, namely SUMO and is compared with the dynamic approach for traffic signal/light control of the present disclosure, with the default option of using fixed-time allocation in SUMO. Recall that the approach of the present disclosure, including the prediction and optimization of a traffic metric with the dynamic online updating capability, determines traffic signal timings with the timings not being fixed as a result of the optimization.

Due to its online prediction and control capabilities, the present disclosure is able to avoid formation of traffic jams and to improve the overall flow (three 3 to thirteen 13 times, depending on the traffic metric) with respect to the SUMO default fixed-time allocation control simulation.

Figure 14 shows some of the preliminary results using two different traffic metrics, namely a default average duration per route length (Fig. 14 left) and a default average duration (Fig. 14 right). The streaming of the traffic data time series is described by a multi-variate ST-ARMA model.

In the above use-case scenario, the traffic model for Shenzhen is used as provided by Shenzhen Police department. In this small-scale scenario, 8 traffic lights systems are considered which control: a highway intersection (4 directions, 5 to 6 lanes per direction), a T-type intersection (3 directions 3 to 4 lanes per direction), and a regular intersection (4 directions, 3 to 4 lanes per direction).

This allows for 30 different routes, on which vehicles are travelling with a certain probability. For the vehicles 16 types are considered, such as a bus, van, passenger and delivery travelling at different speed, acceleration, deceleration, and different levels of driving disciplines, and 20 re routers that teleport certain types of vehicles from a route to another one, based on some probability distribution. There are some other modelling details for the proposed simulation, such as the fact that (i) right turns are always green and (ii) the simulation can run for an indefinitely long time.

Along with the good experimental results, the present disclosure provides serious benefits for a traffic control system that is generic and adaptive, as evident from the comparison results shown in Fig. 14.

The approach and method of the present disclosure is applicable to smart cities infrastructures as well as to any traffic light, respectively any traffic signal. Because the approach does not require any pre-training or manual development to adapt to new layouts, the costs to deploy this solution at scale (to the whole city infrastructure, in any city) is eliminated. This makes the approach of the present disclosure valuable and feasible to support smart city traffic management.

The invention has multiple benefits as it spans from generic approaches for modeling time series prediction up to optimized and dedicated prediction and control on streams.

The core contribution and benefits of the present disclosure is the approach for stream-based prediction and control of traffic flow and city traffic, based on analytical online predictions with adaptive capabilities. This comes from the capability of the system to support real-time adaptation of the scheduling control, based on an up-to-date multi-dimensional view of the environment. Such a system adapts to any type and shape of intersection, works with any traffic metric or characteristic, and it can be deployed at many levels (i.e. central backend, edge/traffic controller).

From the technical point of view, the present disclosure provides a high-implementation benefit as it relies on three efficient and novel stream processing units: A stream processing unit capable to model and fuse (integrate) multi-dimensional time series and to generate predictions based on their correlation structure.

A stream processing unit capable to synchronize the state of predictor models across such compute units, as well as to update the models based on adjacent weighting of compute units.

A stream compute unit with a capability to adapt to any type of intersection and to control the flow without any configuration costs.

• A stream compute unit capable to optimize the scheduling considering constraints.

Overall, the present disclosure provides a generic approach for time series modelling, which is capable to operate on any metrics. Therefore, in the case of traffic optimization, the system can be plugged over the various existing types of traffic sensors. Moreover, using a fast stream processing infrastructure, the disclosure offers a low-latency prediction (i.e. millisecond level) and a fusion for multiple time series as well as a low-latency selection of scheduling policies to maximize a configurable metrics to be optimized. Summarizing, the present disclosure relates to an apparatus and method for traffic control of vehicles at one or more intersections via traffic signals. Each intersection may have a plurality of lanes and the lane may include a traffic signal and a traffic signal timing, defining a state of a lane. For each of the plurality of states of a first intersection, traffic data is acquired and may include traffic data from lanes of intersections neighboring the first intersection. The acquired traffic data are used to predict traffic data at multiple predefined timings of traffic signals for each state. Among the predefined timings, a traffic signal timing is determined for each lane by optimizing a predetermined function of the predicted traffic data of all lanes. The traffic signal timings are then used to control the traffic signals for the one or more lanes.

LIST OF REFERENCE SIGNS

100 Example Scenery

1 10 Intersection

120 Traffic Light Sequence same Timings Fig. 2

200 Example Scenery

210 Intersection

220 Traffic Light Sequence different Timings

230 Multiple-Lane Camera

240 Lane Camera

250 Total Number of Vehicle Counting

260 Intersection Camera

Fig. 3

300 Data Aggregation in finite-sized Window

Fig. 4

400 Example of ML-based Traffic Prediction

410 Four-Lane Intersection

420 Neural Network

Fig. 5

500 Example Embodiment of Data Acquisition and Prediction

Fio. 6

600 Schematic of Traffic Flow Optimizer

610 Traffic Predictor

620 Traffic Optimizer

630 Predictor Updater

Fio. 7

700 Online T raffic Controller Architecture Fio. 8

800 Functional Architecture of T raffic Light Optimization

810 Predictor Model 1

820 Predictor Model 2

Fio. 9

900 State Reduction

Fig. 10

1000 State Selection

Fig. 1 1

1 100 Predictor Update Fig. 12

1200 Online Traffic Control System for one Intersection

Fig. 13

1300 Online Traffic Control System for a Sample City

Fig. 14

1400 Benchmark of T raffic Flow Control System

Claims

1. An apparatus for traffic signal control at a first intersection including a plurality of lanes and a traffic signal for each of the plurality of lanes, wherein the apparatus comprises: a processing circuitry configured to: acquire traffic data for a state among a plurality of states, wherein the state is a combination of one or more lanes belonging to the first intersection and a same traffic signal timing of the traffic signal for the one or more lanes; predict traffic data for each state of the plurality of states based on the acquired traffic data; determine for each state among the plurality of states a first traffic signal timing out of a plurality of predetermined traffic signal timings by optimizing a predetermined function of the predicted traffic data; and control a first traffic signal for each of the plurality of lanes according to the determined first traffic signal timing.

2. The apparatus according to claim 1 , wherein the traffic data is based on a traffic metric with the traffic metric being at least one of a number of vehicles on the lane, an average speed of vehicles on the lane, a vehicle occupation of the lane, a vehicle queue length of the lane, and/or an average waiting time of vehicles on a lane.

3. The apparatus according to claim 1 or 2, wherein the processing circuitry is further configured to optimize the traffic data according to a constraint, including at least one of a maximum cycle time of the traffic signal timing, minimum green time per lane, and yellow time.

4. The apparatus according to claim 3, wherein the constraint is predefined or selected based on the traffic metric.

5. The apparatus according to claim 4, wherein the plurality of states includes states of a plurality of lanes of a second intersection spatially neighboring the first intersection.

6. The apparatus according to claim 5, wherein the acquiring of the traffic data for a state includes traffic data of a state of a lane among the plurality of lanes of the second intersection.

7. The apparatus according to any one of claims 1 to 6, wherein the traffic data acquisition is further based on a traffic data history and the processing circuitry is configured to: acquire traffic data from the traffic data history acquired at a plurality of time points earlier than a time point of the traffic data acquisition.

8. The apparatus according to any one of claims 1 to 7, wherein the prediction of the traffic data is based on a parametric prediction model for the state and the processing circuitry is configured to: update a parameter of the prediction model in accordance to the acquired traffic data; - the predicted traffic data; and/or the spatially neighboring second intersection.

9. The apparatus according to claim 8, wherein the processing circuitry is further configured to update the parameter of the prediction model of a first state based on a second state selected among the plurality of states.

10. The apparatus according to claim 9, wherein the selection of the second state is based on a selection policy referring to a state for which traffic data has been acquired, a state for which traffic data has been predicted, and/or a state whose traffic data has been obtained by transforming a traffic metric of a different state.

1 1. A method for traffic signal control at a first intersection including a plurality of lanes and a traffic signal for each of the plurality of lanes, wherein the method comprises the steps of: acquiring traffic data for a state among a plurality of states, wherein the state is a combination of one or more lanes belonging to the first intersection and the traffic signal for the one or more lanes being same over a traffic signal timing out of a plurality of predetermined traffic signal timings; predicting traffic data for each state of the plurality of states based on the acquired traffic data; determining for each state among the plurality of states a first traffic signal timing out of the plurality of predetermined traffic signal timings by optimizing a predetermined function of the predicted traffic data; and controlling a first traffic signal for each of the plurality of lanes according to the determined first traffic signal timing.

12. A computer-readable non-transitory medium storing a program, including instructions which when executed on a processor cause the processor to perform the method according to claim 1 1 .