CN105930855A - Vehicle detection method based on deep convolution neural network - Google Patents

Abstract

The invention discloses a vehicle detection method based on a deep convolution neural network. The vehicle detection method comprises the following steps: acquiring road surface pictures in real time through a camera, and normalizing sizes of the pictures; selecting a threshold to segment a road surface vehicle shadow area; determining a vehicle candidate area in the road surface vehicle shadow area; and training the convolution neural network, recognizing the vehicle candidate area by using the trained convolution neural network, and outputting a detection result. According to the vehicle detection method, the convolution neural network is adopted to perform verification, so that the timeliness and accuracy of a whole system are improved.

Description

A kind of vehicle checking method based on degree of depth convolutional Neural net

Technical field

The present invention relates to road safety field, be specifically related to a kind of vehicle checking method based on degree of depth convolutional Neural net.

Background technology

Undoubtedly, along with socioeconomic development, the requirement of the vehicles is the most constantly improved by people, automobile Occur bringing huge facility to the life of people undoubtedly, change the life style of people, improve the living standard of people. The automobile pollution of China is very big at present, motor vehicles more than 200,000,000, nearly 100,000,000 of automobile, light production and marketing in 2010 more than 17,000,000. But being as the increase of automobile pollution, the safety problem that the vehicles are brought easily does not allows people to be ignored, frequently yet Vehicle accident, heavy casualties and huge property loss to make motor traffic safety problem be increasingly becoming of concern Focus.According to ASSOCIATE STATISTICS, in annual worldwide road traffic accident, about 10,000,000 people are injured, and wherein severe injury is about 3000000 people, dead 400,000 people, the economic loss that these accidents cause accounts for the 13% of world GDP.By lot of accident is divided Analysis, the accident caused due to the reason of non-vehicle own reaches 70%: cause of accident mainly with the active safety of driver because have Close.Traffic safety is to hinder society and a big problem of economic development, and all deeply hurting in countries in the world and area, is therefore subject to To countries in the world government and the concern of society.Although having worked out many passive security measures to reduce accident After casualties, but cause vehicle accident basic reason solved the most at all.Research shows, as long as there being collision danger The 0.5s forward direction driver of danger sounds a warning, it is possible to avoid the car accident that knocks into the back of 60%, the thing of head-on colliding of 30% Therefore and 50% road surface related accidents；If there being " early warning " time of 1s, the accident of 90% can be avoided to occur.

Summary of the invention

For the deficiency overcoming prior art to exist, the present invention provides a kind of vehicle detection based on degree of depth convolutional Neural net Method.

The present invention adopts the following technical scheme that

A kind of vehicle checking method based on degree of depth convolutional Neural net, including:

Photographic head Real-time Collection road surface picture, is normalized picture size；

It is partitioned into road vehicles shadow region by selected threshold；

In road vehicles shadow region, determine vehicle candidate region；

Convolutional neural networks is trained, then uses the convolutional neural networks identification vehicle candidate region trained, Output detections result.

The described convolutional neural networks trained specifically trains the process to be:

S1 utilizes camera collection vehicle sample information, described vehicle sample information to include positive samples pictures and negative sample figure Sheet, carries out Image semantic classification；

Pretreated vehicle pictures as the input of degree of depth convolutional Neural net, is trained the weights in neutral net by S2.

Described photographic head is specially monocular cam.

Described Image semantic classification includes gray processing and mean filter.

Described S2 use back-propagation algorithm degree of depth convolution net is trained.

Described in road vehicles shadow region, determine that vehicle candidate region specifically uses calculating vehicle length-width ratio, according to Shade width delimit region to be identified.

It is partitioned into road vehicles shade by selected threshold, particularly as follows:

By pretreated picture, sample area gray value statistics grey value profile；

Calculate average m and variances sigma²；

If σ²＞ threshold value S, then delete the gray value more than average, average statistical and variance again；

Calculate threshold value T=m-3 σ, use this threshold binarization picture, obtain road vehicles shadow region.

Beneficial effects of the present invention:

The present invention identifies vehicle by convolutional neural networks, it is to avoid traditional recognition method uses the artificial feature extracted Limitation, convolutional neural networks by representing the best features of object from great amount of samples learning, and therefore accuracy rate is high.

Real-time in order to avoid using sliding window detection to bring reduces problem, and the present invention uses shadow Detection from figure The position that quick location vehicle exists, then verifies with convolutional neural networks, thus improves the real-time of whole system with accurate Property.

Accompanying drawing explanation

Fig. 1 is the workflow schematic diagram of the present invention；

Fig. 2 is the schematic flow sheet training neutral net in Fig. 1；

Fig. 3 is the schematic flow sheet of selected threshold；

Fig. 4 determines that the schematic flow sheet of vehicle candidate region.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not It is limited to this.

Embodiment

As Figure 1-Figure 4, a kind of vehicle checking method based on degree of depth convolutional Neural net, including:

Photographic head Real-time Collection road surface picture, carries out pretreatment and normalization to its size, and described pretreatment includes gray scale Changing and mean filter, described photographic head is monocular cam；

Vehicle shadow region, road surface is gone out by choosing suitable Threshold segmentation；

Owing to general road image 1/3 is the irrelevant information such as sky, mountain, therefore we only need to detect below image 1/3 Region, in general, shade is that illumination is caused by vehicle shelter, and therefore the gray value of general shaded side is lower than road surface, The most only need to obtain the normal gray value on road surface, i.e. can be partitioned into bottom shadow.

In order to prevent the impact of the land marking such as deceleration strip, word, herein by data and the historical data phase of current statistic In conjunction with method choose suitable threshold value, judge whether road surface exists deceleration strip, literary composition by assembly average m and variances sigma The land markings such as word.

The most as shown in Figure 3:

By pretreated picture, sample area gray value statistics grey value profile；

Calculate average m and variances sigma²；

If σ²＞ threshold value S, then delete the gray value more than average, recalculate average and variance, enter next step, as Really σ²＜ threshold value S, then be directly entered next step；

Calculate threshold value T=m-3 σ, use this threshold binarization picture, obtain road vehicles shadow region.

As shown in Figure 4, in road vehicles shadow region, determine vehicle candidate region；

For different vehicles, its length-width ratio is different, by the vehicle length-width ratio that statistics is common, according to shade width Degree, delimit candidate region to be identified.

Generally in road picture, it is 3:2 by the ratio of shade length and overall height, therefore just can be by vehicle by this ratio Outline completely from figure.

Convolutional neural networks is trained, then uses the convolutional neural networks identification vehicle candidate region trained, Output detections result.

It is the input of network owing to convolutional Neural net can accept gray scale picture initial data, therefore chooses in candidate region After, the output obtaining network with this as the input of network, determine whether vehicle by output.

The process that described convolutional neural networks is specifically trained is:

S1 utilizes camera collection vehicle sample information, described vehicle sample information to include positive samples pictures and negative sample figure Sheet, carries out Image semantic classification；Obtaining pretreated vehicle as positive samples pictures, described positive sample is the figure comprising vehicle Sheet, does not comprise the negative sample picture of vehicle, and positive sample size is more than 2000, and negative sample quantity is more than 4000, generally Negative sample quantity is the twice of positive sample size；

Pretreated vehicle pictures as the input of degree of depth convolutional Neural net, is trained the weights in neutral net by S2, Specifically use back propagation:

First initialize network weight, can be randomly derived, be then iterated training, specifically comprise the following steps that

First stage, propagated forward, the training of neutral net, its purpose is through existing information and constantly adjusts net The parameter of network self makes the output of network become closer to actual output, i.e. error is more and more less.For n sample, seek Look for the minima of Formula Solution, first calculate the error of sample output

$E^N(W,b;x,y)=\frac{1}{2}\underset{n=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{c}{\Σ}}{(t_k^n-y_k^n)}^2$

1. convolutional layer: assuming that m layer is convolutional layer, m+1 layer is down-sampling layer, then the calculating of m layer jth characteristic pattern Formula is as follows:

$x_j^m=f(\underset{i\∈M_j}{\Σ}{x}_i^{m-1}*w_{ij}^m+b_j^m)$

2. down-sampling layer: choose maximum sampling, i.e. choose the maximum feature generation as one's respective area of a region class Table.

The propagated forward of convolutional neural networks is through continuous convolution and down-sampling, finally by the full articulamentum of multilamellar Obtain final output

Second stage: back propagation, updates weights W and b.

Weights W and d updated in network by gradient descent method is updated by column, the following is the most former of gradient descent method Reason:

$W_{ij}^m=W_{ij}^m-\mathrm{\α}\frac{\∂}{\∂W_{ij}^m}E(W,b)$

$b_i^m=b_i^m-\mathrm{\α}\frac{\∂}{\∂b_i^m}E(W,b)$

Repeatedly perform the step in above-mentioned first and second stage, obtain W and b of convergence.

Described Image semantic classification includes gray processing and mean filter, particularly as follows:

Image gray processing, the picture that photographic head is gathered is generally colour picture, and we have only to the half-tone information of image, The gray value of image is calculated hence with below equation

Gray=0.11B+0.59G+0.3R

Mean filter, there is certain noise in the picture captured by photographic head, therefore reason mean filter carrys out smooth noise, Meansigma methods mainly by certain pixel neighboring pixel gets to the effect of smooth noise.

Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention are not by described embodiment Limit, the change made under other any spirit without departing from the present invention and principle, modify, substitute, combine, simplify, All should be the substitute mode of equivalence, within being included in protection scope of the present invention.

Claims (7)

1. a vehicle checking method based on degree of depth convolutional Neural net, it is characterised in that comprise the steps:

Photographic head Real-time Collection road surface picture, is normalized picture size；

It is partitioned into road vehicles shadow region by selected threshold；

In road vehicles shadow region, determine vehicle candidate region；

Convolutional neural networks is trained, then uses the convolutional neural networks identification vehicle candidate region trained, output Testing result.

Vehicle checking method the most according to claim 1, it is characterised in that the described convolutional neural networks trained is concrete Training process is:

S1 utilizes camera collection vehicle sample information, described vehicle sample information to include positive samples pictures and negative sample picture, Carry out Image semantic classification；

Pretreated vehicle pictures as the input of degree of depth convolutional Neural net, is trained the weights in neutral net by S2.

Vehicle checking method the most according to claim 1 and 2, it is characterised in that described photographic head is specially monocular shooting Head.

Vehicle checking method the most according to claim 2, it is characterised in that described Image semantic classification includes gray processing and all Value filtering.

Vehicle checking method the most according to claim 2, it is characterised in that use back-propagation algorithm to deeply in described S2 Degree convolution net is trained.

Vehicle checking method the most according to claim 1, it is characterised in that described in road vehicles shadow region, determines Vehicle candidate region specifically uses calculating vehicle length-width ratio, delimits region to be identified according to shade width.

Vehicle checking method the most according to claim 1, it is characterised in that be partitioned into road vehicles by selected threshold cloudy Territory, shadow zone, particularly as follows:

By pretreated picture, sample area gray value statistics grey value profile；

Calculate average m and variances sigma²；

If σ²＞ threshold value S, then delete the gray value more than average, and average statistical and variance again, subsequently into next step；As Really σ²＜ threshold value S, then be directly entered next step；

Calculate threshold value T=m-3 σ, use this threshold binarization picture, obtain road vehicles shadow region.