CN115456064B - Object classification method based on point cloud and related equipment - Google Patents
Object classification method based on point cloud and related equipment Download PDFInfo
- Publication number
- CN115456064B CN115456064B CN202211076689.7A CN202211076689A CN115456064B CN 115456064 B CN115456064 B CN 115456064B CN 202211076689 A CN202211076689 A CN 202211076689A CN 115456064 B CN115456064 B CN 115456064B
- Authority
- CN
- China
- Prior art keywords
- global
- attention mechanism
- global attention
- feature matrix
- determining
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 91
- 230000007246 mechanism Effects 0.000 claims abstract description 166
- 239000011159 matrix material Substances 0.000 claims abstract description 152
- 238000000605 extraction Methods 0.000 claims description 47
- 230000009466 transformation Effects 0.000 claims description 47
- 238000012545 processing Methods 0.000 claims description 28
- 238000003860 storage Methods 0.000 claims description 15
- 238000011176 pooling Methods 0.000 claims description 14
- 239000000758 substrate Substances 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 10
- 238000010586 diagram Methods 0.000 description 13
- 238000004590 computer program Methods 0.000 description 11
- 239000013598 vector Substances 0.000 description 11
- 238000004364 calculation method Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 9
- 230000008901 benefit Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 5
- 238000009826 distribution Methods 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 230000006872 improvement Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000003190 augmentative effect Effects 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
- G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an object classification method based on point cloud and related equipment, relates to the field of point cloud, and mainly aims to solve the problem that classification accuracy and stability are difficult to consider when object classification is carried out based on point cloud. The method comprises the following steps: determining Ji Dian cloud coordinate data of a pair of target objects; determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism; global features of the target object are determined based on the high-level feature matrix to determine a classification result. The method is used for the object classification process based on the point cloud.
Description
Technical Field
The invention relates to the field of point clouds, in particular to an object classification method based on the point clouds and related equipment.
Background
Object classification is a classical problem in visual computing and pattern recognition. With the development of deep neural network technology, the object classification performance is rapid, and the method has stronger application potential in robots, automatic driving and augmented reality. Common object representation methods include images and point clouds. Because of the natural ordering, uniformity and regularity of image structures, deep neural network technology has first been successful in classifying objects with images as input. Compared with image input, although the three-dimensional point cloud has the advantages of richer spatial information and less influence of illumination change, the natural disorder, non-uniformity and irregularity of the three-dimensional point cloud make designing a neural network feature extraction and classification method directly taking the three-dimensional point cloud as input challenging.
The current common classification methods are: global feature-based methods, local feature-based methods, and neighborhood feature-based methods. Although the method based on global features has very strong stability to the change of the point cloud density caused by the reasons of the distance between the target capture and the like because the features of the points are not influenced by the distribution of surrounding points, the method has the defect of poor classification precision; the method based on the local characteristics and the method based on the neighborhood characteristics consider the local characteristics and the neighborhood characteristics of the point cloud, so that the performance of the method is affected by the local missing and distribution change of the point cloud. Therefore, the technical problem that the classification precision and the stability are difficult to be compatible still exists in the prior art.
Disclosure of Invention
In view of the above problems, the present invention provides a method and related equipment for classifying objects based on point clouds, and is mainly aimed at solving the problem that classification accuracy and stability are difficult to be compatible when classifying objects based on point clouds.
To solve at least one of the above technical problems, in a first aspect, the present invention provides a method for classifying objects based on a point cloud, the method comprising:
determining Ji Dian cloud coordinate data of a pair of target objects;
determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism;
And determining global features of the target object based on the high-level feature matrix to determine a classification result.
Optionally, the method further comprises:
determining point cloud coordinate data based on the target object;
the point cloud coordinate data is spatially transformed based on a spatial transformation network to determine the pair Ji Dian cloud coordinate data.
Optionally, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
a high-level feature matrix is determined based on the pair Ji Dian cloud coordinate data and the target global feature extraction model based on the global attention mechanism.
Optionally, the target global feature extraction model based on the global attention mechanism includes: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
Alternatively to this, the method may comprise,
the cascade global attention mechanism is formed by cascading a plurality of global attention mechanisms,
the multi-layer perceptron network is used for extracting the characteristics of the point cloud data,
The feature transformation network is used for aligning the features of the point cloud data.
Optionally, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
performing multi-layer perceptron network processing based on a global attention mechanism on the Ji Dian cloud coordinate data to obtain a low-layer feature matrix;
performing feature transformation network processing based on a cascade global attention mechanism on the low-level feature matrix to acquire an aligned low-level feature matrix;
and performing multi-layer perceptron network processing based on a cascade global attention mechanism on the aligned low-layer feature matrix to acquire a high-layer feature matrix.
Optionally, the determining the global feature of the target object based on the high-level feature matrix to determine a classification result includes:
carrying out maximum pooling treatment on the high-level feature matrix to obtain global features;
and carrying out full-connection network processing on the global features to classify the target objects.
In a second aspect, an embodiment of the present invention further provides an object classification device based on a point cloud, including:
a first determining unit for determining a pair Ji Dian cloud coordinate data of the target object;
A second determining unit, configured to determine a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism;
and a third determining unit, configured to determine global features of the target object based on the high-level feature matrix to determine a classification result.
In order to achieve the above object, according to a third aspect of the present invention, there is provided a computer-readable storage medium including a stored program, wherein the steps of the above-described point cloud-based object classification method are implemented when the program is executed by a processor.
In order to achieve the above object, according to a fourth aspect of the present invention, there is provided an electronic device including at least one processor, and at least one memory connected to the processor; the processor is used for calling the program instructions in the memory and executing the object classification method based on the point cloud.
By means of the technical scheme, the invention provides an object classification method based on point cloud and related equipment. For the problem that classification accuracy and stability are difficult to consider when object classification is carried out based on point clouds, the method determines Ji Dian cloud coordinate data of a target object; determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism; and determining global features of the target object based on the high-level feature matrix to determine a classification result. In the scheme, as three key networks of the point cloud target global feature extraction model are redesigned, and different global attention mechanisms are organically fused with the original network in the new network, each point in the point cloud can fully utilize the features of all points in each key stage of feature extraction, so that the classification precision is improved, meanwhile, the classification precision is not related to the division of local point cloud or the calculation of the neighborhood of the point in each stage of feature extraction, the classification stability is ensured, and the technical problem that the classification precision and the stability are difficult to consider in the prior art is solved.
Accordingly, the object classification device, the device and the computer readable storage medium based on the point cloud provided by the embodiment of the invention also have the technical effects.
The foregoing description is only an overview of the present invention, and is intended to be implemented in accordance with the teachings of the present invention in order that the same may be more clearly understood and to make the same and other objects, features and advantages of the present invention more readily apparent.
Drawings
Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. Also, like reference numerals are used to designate like parts throughout the figures. In the drawings:
fig. 1 shows a schematic flow chart of an object classification method based on point cloud according to an embodiment of the present invention;
fig. 2 is a schematic diagram of a network structure of an object classification method based on point cloud according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a multi-layer perceptron network based on a global attention mechanism according to an embodiment of the present invention;
FIG. 4 illustrates a schematic diagram of a cascaded global attention mechanism provided by an embodiment of the present invention;
FIG. 5 shows a schematic diagram of a feature transformation network based on a cascaded global attention mechanism according to an embodiment of the present invention;
FIG. 6 illustrates a schematic diagram of a multi-layer perceptron network based on a cascaded global attention mechanism provided by an embodiment of the present invention;
fig. 7 is a schematic block diagram illustrating the composition of an object classification device based on a point cloud according to an embodiment of the present invention;
fig. 8 shows a schematic block diagram of a point cloud-based object classification electronic device according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
In order to solve the problem that classification accuracy and stability are difficult to consider when classifying objects based on point clouds, an embodiment of the present invention provides a method for classifying objects based on point clouds, as shown in fig. 1, the method includes:
S101, determining the pair Ji Dian cloud coordinate data of a target object;
illustratively, the point data set of the product appearance surface obtained by the measuring instrument in reverse engineering is also referred to as point cloud. The point cloud is a massive point set expressing target space distribution and target surface characteristics under the same space reference system, and is often directly obtained by measurement. Each point corresponds to a measuring point and is not subjected to other processing means, so that the maximum information quantity is contained. The point cloud obtained according to the laser measurement principle comprises three-dimensional coordinates and laser reflection intensity. The point cloud obtained according to the photogrammetry principle comprises three-dimensional coordinates and color information. And combining laser measurement and photogrammetry principles to obtain a point cloud, wherein the point cloud comprises three-dimensional coordinates, laser reflection intensity and color information. After the spatial coordinates of each sample point on the object surface are obtained, a set of points, called a "point cloud", is obtained. The method comprises the steps of firstly obtaining Ji Dian cloud coordinate data of a target object to be classified.
S102, determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism;
by way of example, the method introduces a global attention mechanism technology into three-dimensional point cloud classification, and by redesigning three key networks of a point cloud target global feature extraction model, and organically fusing different global attention mechanism modules with an original network architecture in a new network, a high-level feature matrix is determined.
S103, determining global features of the target object based on the high-level feature matrix to determine a classification result.
By means of the method, the characteristics of all points can be fully utilized in each key stage of characteristic extraction of each point in the point cloud, and division of the local point cloud or calculation of the neighborhood of the point is not involved in each stage of characteristic extraction, so that the technical problem that classification accuracy and stability are difficult to consider in the prior art is solved.
Illustratively, the proper nouns in this solution are illustrated as follows: the Point Cloud is Point Cloud, the downsampling is Down Sampling, the spatial transformation network is Spatial Transformer Network, the three-dimensional coordinate matrix is 3DCoordinates Matrix, the Aligned three-dimensional coordinate matrix is Aligned 3D Coordinates Matrix, the Feature extraction model is Feature Extraction Model, the Global attention mechanism is Global Attention Mechanism, the cascade Global attention mechanism is Cascaded Global Attention Mechanism, the Multi-layer perceptron network is Multi-Layer Perceptron Network, the Feature transformation network is Feature Transformer Network, the Low-layer Feature matrix is Low-Level Feature Matrix, the High-layer Feature matrix is High-Level Feature Matrix, the Aligned Low-layer Feature matrix is Aligned Low-Level Feature Matrix, the maximum Pooling is Max-Pooling, the Fully connected network is Fully-Connected Network, the Global Feature is Global Feature, the classification Vector is Class Vector, the classification precision is Classification Accuracy, and the Stability is Stability. .
By means of the technical scheme, the object classification method based on the point cloud is provided. For the problem that classification accuracy and stability are difficult to consider when object classification is carried out based on point clouds, the method determines Ji Dian cloud coordinate data of a target object; determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism; and determining global features of the target object based on the high-level feature matrix to determine a classification result. In the scheme, as three key networks of the point cloud target global feature extraction model are redesigned, and different global attention mechanisms are organically fused with the original network in the new network, each point in the point cloud can fully utilize the features of all points in each key stage of feature extraction, so that the classification precision is improved, and the division of local point cloud or the calculation of the neighborhood of the point is not involved in each stage of feature extraction, the classification stability is ensured, and the technical problem that the classification precision and the stability are difficult to consider in the prior art is solved.
In one embodiment, the method further comprises:
determining point cloud coordinate data based on the target object;
The point cloud coordinate data is spatially transformed based on a spatial transformation network to determine the pair Ji Dian cloud coordinate data.
For convenience of description, let p= [ P ] 1 ,p 2 ,K,p N ] T Three-dimensional coordinate matrix representing downsampled n×3-dimensional input point cloud, i.e., the above-described point cloud coordinate data, where p i And (3) representing three-dimensional coordinate vectors of the ith point in the three-dimensional coordinate matrix of the input point cloud after downsampling, wherein N represents the number of points in the input point cloud after downsampling, and T represents the transposition of the matrix. Let C denote the output classification vector of dimension C, where C denotes the number of classes. In the process of classifying the point cloud object, the input point cloud P first obtains an aligned n×3 dimensional coordinate matrix, that is, the pair Ji Dian cloud coordinate data through a spatial transformation network.
In one embodiment, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
a high-level feature matrix is determined based on the pair Ji Dian cloud coordinate data and the target global feature extraction model based on the global attention mechanism.
By way of example, the method redesigns the global feature extraction model of the point cloud target, and in the new model, different global attention mechanisms are organically integrated with the original global feature extraction framework, so that each point in the point cloud can fully utilize the features of all points under the assistance of the global attention mechanism in each key stage of feature extraction, and the classification precision of the point cloud object is greatly improved. Meanwhile, each stage of feature extraction does not involve the division of local point clouds or the calculation of the neighborhood of the points, so that the influence of local loss and distribution change of the point clouds is avoided, and the stability of super-strong classification precision can be maintained under the extreme condition of sharp reduction of the number of the point clouds.
In one embodiment, the target global feature extraction model based on the global attention mechanism includes: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
Illustratively, the method redesigns three key networks of the point cloud target global feature extraction model, namely a multi-layer perceptron network based on a global attention mechanism for extracting low-level features, a feature transformation network based on a cascade global attention mechanism for aligning low-level features and a multi-layer perceptron network based on a cascade global attention mechanism for extracting high-level features. In the new network, different global attention mechanisms are organically integrated with the original global feature extraction architecture, so that the features of all points can be fully utilized by each point in the point cloud in each key stage of feature extraction, and the classification precision of the point cloud objects is greatly improved.
In one embodiment of the present invention, in one embodiment,
the cascade global attention mechanism is formed by cascading a plurality of global attention mechanisms,
the multi-layer perceptron network is used for extracting the characteristics of the point cloud data,
the feature transformation network is used for aligning the features of the point cloud data.
For example, two subsequent key networks, a feature transformation network based on a cascade global attention mechanism and a multi-layer perceptron network based on a cascade global attention mechanism, are used for the cascade global attention mechanism, and the main function of the mechanism is to obtain global features under different attention concentration degrees. As shown in fig. 4, the cascaded global attention mechanism is designed as follows. The mechanism is formed by cascading m global attention mechanisms with the same structure in sequence. In the process of extracting global features under different attention concentration degrees, an N x D-dimensional input feature matrix firstly passes through a global attention mechanism 1 to obtain a first N x D-dimensional feature matrix, the feature matrix passes through a global attention mechanism 2 to obtain a second N x D-dimensional feature matrix, and the like until passing through a global attention mechanism m, wherein D represents the dimension of an input feature vector at the midpoint of a point cloud. And finally, sequentially splicing the obtained m N multiplied by D dimensional matrixes to form a final N multiplied by mD dimensional characteristic. As shown in fig. 4, as the feature matrix continuously passes through the global attention mechanism, the focusing degree of the features is continuously deepened, and the features with different focusing degrees are cascaded together, so that the obtained final features can more accurately represent the features with different scales in the object point cloud. Therefore, compared with a single global attention mechanism, the features obtained by cascading the global attention mechanism have stronger resolution, and are beneficial to further improving the overall classification precision.
In one embodiment, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
performing multi-layer perceptron network processing based on a global attention mechanism on the Ji Dian cloud coordinate data to obtain a low-layer feature matrix;
performing feature transformation network processing based on a cascade global attention mechanism on the low-level feature matrix to acquire an aligned low-level feature matrix;
and performing multi-layer perceptron network processing based on a cascade global attention mechanism on the aligned low-layer feature matrix to acquire a high-layer feature matrix.
Illustratively, the function of the multi-layer perceptron network based on the global attention mechanism is to extract low-level features from the pair Ji Dianyun coordinates. As shown in fig. 3, the network is composed of a multi-layer perceptron and a global attention mechanism. In the process of extracting low-level features, aligning an N multiplied by 3 dimensional point cloud coordinate matrix, and firstly obtaining N multiplied by D through a multi-layer perceptron L A dimensional feature matrix, wherein D L Is the dimension of the low-level features in the point cloud. The feature matrix and NxD obtained by the global attention mechanism module L Adding the dimension feature matrix to obtain NxD L And (5) maintaining a low-level feature matrix. As shown in fig. 3, compared with the multi-layer perceptron network for extracting low-layer features in the point cloud object classification method based on global features, the method adds a global attention mechanism module. In the process of extracting low-level characteristics, a global note is superimposed on the characteristics output by the original multi-layer perceptron The characteristics of the force module output. The network design effectively enhances the identification force of low-level features and is beneficial to the improvement of the overall classification precision and stability due to the fact that the global attention mechanism focuses on important features in a global range and improves the stability.
Illustratively, the feature transformation network based on the cascading global attention mechanism functions to align the extracted low-level features. As shown in fig. 5, the network is composed of three multi-tier perceptrons, a cascaded global attention mechanism and a max pooling tier. In the process of aligning low-level features, N x D is input L The low-level feature matrix is first passed through the multi-layer perceptron 1 to obtain NxD 1 A dimensional feature matrix, wherein D 1 The dimension of the feature of the point output by the multi-layer perceptron 1. Then obtaining Nxm by cascading the global attention mechanism modules 1 D 1 A dimensional feature matrix, wherein m 1 Representing the number of cascades of global attention mechanism modules for the feature transformation network. Then through the multi-layer perceptron 2, N x D is obtained 2 A dimensional feature matrix, wherein D 2 The dimension of the feature of the point output by the multi-layer perceptron 2. And then obtaining 1 xD through a maximum pooling layer 2 And (5) a dimension vector. Finally, obtaining D through a multi-layer perceptron 3 L ×D L And (5) a dimensional feature transformation matrix. To be input NxD L Multiplying the matrix by the low-level feature matrix to obtain an aligned NxD L And (5) maintaining a low-level feature matrix. As shown in fig. 5, compared with a feature transformation network for aligning low-level features in the point cloud object classification method based on global features, the network adds a cascading global attention mechanism in the feature transformation matrix solving part. The design can more comprehensively and accurately align the object point cloud characteristics under each scale due to the characteristics of multi-level focusing important characteristics of a cascade global attention mechanism in a global range, and is beneficial to improving the overall classification precision.
Illustratively, the function of the multi-layer perceptron network based on cascaded global attention mechanisms is to extract high-level features from the aligned low-level features. As shown in fig. 6, a multi-layer perceptron network based on cascaded global attention mechanisms is designed as follows, which network is composed ofTwo multi-layer perceptrons and a cascaded global attention mechanism. In the process of extracting high-level features, the alignment low-level feature matrix is firstly passed through a multi-layer perceptron 1 to obtain NxD 3 A dimensional feature matrix, wherein D 3 The dimension of the feature of the point output by the multi-layer perceptron 1. Then N x m is obtained by cascading global attention mechanisms 2 D 3 A dimensional feature matrix, wherein m 2 Representing the number of cascades of global attention mechanism modules for a multi-tier perceptron network. Finally, obtaining NxD through the multi-layer perceptron 2 H A high-level feature matrix, D H Is the dimension of the high-level features in the point cloud. As shown in fig. 6, compared with the multi-layer perceptron network for extracting high-level features in the point cloud object classification method based on global features, the network is newly added with a cascade global attention mechanism. The network design comprehensively enhances the identification force of high-level features and is beneficial to the improvement of the overall classification precision and stability due to the characteristics of multi-scale focusing important features and multi-level stability promotion of a cascade global attention mechanism in a global range.
In one embodiment, the determining the global feature of the target object based on the high-level feature matrix to determine the classification result includes:
carrying out maximum pooling treatment on the high-level feature matrix to obtain global features;
and carrying out full-connection network processing on the global features to classify the target objects.
Exemplary, N X D is obtained H After the high-level feature matrix of the dimension, obtaining 1 xD through a maximum pooling layer H And finally, obtaining an output classification vector C of the dimension C through a fully connected network to classify the target object.
Illustratively, as shown in FIG. 2, the overall network framework is designed as follows. The framework sequentially comprises a space transformation network, a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascade global attention mechanism, a multi-layer perceptron network based on the cascade global attention mechanism, a maximum pooling layer and a full connection network. At the pointIn the cloud object classification process, an input point cloud P firstly obtains an aligned point cloud, namely an aligned N multiplied by 3 dimensional coordinate matrix through a space transformation network, and then obtains N multiplied by D through a multi-layer perceptron network based on a global attention mechanism L The low-level feature matrix of the dimension is then passed through a feature transformation network based on a cascade global attention mechanism to obtain an nxd L The alignment low-level characteristic matrix of the dimension is obtained by a multi-layer perceptron network based on a cascade global attention mechanism H The high-level characteristic matrix of the dimension is obtained by the maximum pooling layer to obtain 1 xD H And finally obtaining a C-dimensional output classification vector C through the global feature of the dimension and the full connection network.
Because the method redesigns three key networks of the point cloud object classification method based on the global features, namely a multi-layer perceptron network for extracting low-level features, a feature transformation network for aligning the low-level features and a multi-layer perceptron network for extracting high-level features. In the new network, different global attention mechanisms are organically integrated with the original network architecture, so that the characteristics of all points can be fully utilized at each key stage of characteristic extraction by each point in the point cloud, and the classification accuracy of the point cloud objects is greatly improved. And the division of local point clouds or the calculation of the neighborhood of the points are not involved in each stage of feature extraction, so that the method is not influenced by the local deletion and distribution change of the point clouds, and can still maintain the stability of super-strong classification precision under the extreme condition of sharp reduction of the number of the point clouds, unlike the existing method based on the local features and the method based on the neighborhood features. In the three global attention mechanism modules used in the method, the feature calculation of each point is related to all points in the point cloud, and the calculation cost is proportional to the square of the number of points in the input point cloud. The down sampling in the point cloud preprocessing can effectively control the number of points in the input point cloud, so that compared with the point cloud object classification method based on global features, the three global attention mechanism modules have smaller newly-added calculation cost. Therefore, the method well maintains the advantages of moderate parameters and high calculation efficiency of the point cloud object classification method based on the global features. Meanwhile, the three global attention mechanism modules have the characteristic of improving the feature resolution, so that the overall classification accuracy is greatly improved.
As an implementation of the method, specific implementation steps may be:
(1) Performing space transformation on the down-sampled input point cloud coordinate matrix P to obtain an aligned N multiplied by 3 dimensional point cloud coordinate matrix;
(2) Obtaining NxD through a multi-layer perceptron network based on global attention mechanisms L Maintaining a low-level feature matrix;
(2.1) inputting the aligned N multiplied by 3 dimensional point cloud coordinate matrix obtained in the step (1) into a multi-layer perceptron based on a global attention mechanism multi-layer perceptron network to obtain N multiplied by D of an object point cloud L A dimensional feature matrix;
(2.2) subjecting the above-mentioned NxD L The dimension feature matrix inputs the global attention mechanism based on the global attention mechanism multi-layer perceptron network to obtain NxD L A dimensional feature matrix;
(2.3) N.times.D obtained in step (2.1) L The dimension characteristic matrix and the NxD obtained in the step (2.2) L Adding the dimension feature matrix to obtain NxD L And (5) maintaining a low-level feature matrix.
(3) Aligned nxd through feature transformation network based on cascaded global attention mechanism L Maintaining low-level features;
(3.1) N X D obtained in the step (2) L The first multi-layer perceptron of the feature transformation network is input by the feature matrix of the low-level dimension to obtain NxD 1 A dimensional feature matrix;
(3.2) subjecting the above-mentioned NxD 1 Cascading global attention mechanism of feature transformation network is input into dimension feature matrix to obtain Nxm 1 D 1 A dimensional feature matrix;
(3.2.1) N.times.D obtained in step (3.1) 1 The dimension characteristic matrix is input to a first global attention mechanism module of the cascade global attention mechanism to obtain N multiplied by D output by the first module 1 A dimensional feature matrix;
(3.2.2) N X D which outputs the first module 1 The dimensional feature matrix is input to a second global attention mechanism module,obtaining the NxD output from the second module 1 A dimensional feature matrix, and so on, until the mth is obtained 1 NxD output by each global attention mechanism module 1 A dimensional feature matrix;
(3.2.3) the m obtained in steps (3.2.1) and (3.2.2) 1 NxD output by each global attention mechanism module 1 The dimension feature matrix is connected end to end, and N multiplied by m output by the cascade global attention mechanism module in the feature transformation network based on the cascade global attention mechanism is obtained 1 D 1 And (5) a dimensional feature matrix.
(3.3) mixing the above-mentioned NXm 1 D 1 The dimension characteristic matrix is input into a second multi-layer perceptron of the characteristic transformation network to obtain NxD 2 A dimensional feature matrix;
(3.4) the above-mentioned NxD 2 The dimension characteristic matrix is input into the maximum pooling layer of the characteristic transformation network to obtain 1 xD 2 A dimension vector;
(3.5) the above 1 XD 2 The third multi-layer perceptron of the dimension vector input feature transformation network obtains D L ×D L A dimensional feature transformation matrix;
(3.6) N X D obtained in the step (2) L Dimension low-level feature matrix and D L ×D L Multiplying the dimension characteristic transformation matrix to obtain aligned NxD L And (5) maintaining a low-level feature matrix.
(4) Obtaining nxd through a multi-layer perceptron network based on cascaded global attention mechanisms H High-level features are maintained;
(4.1) alignment NxD obtained in step (3) L The low-level feature matrix is input into a first multi-layer perceptron of a multi-layer perceptron network based on a cascade global attention mechanism to obtain NxD 3 A dimensional feature matrix;
(4.2) subjecting the above-mentioned NxD 3 The dimension feature matrix is input into a cascade global attention mechanism of a multi-layer perceptron network based on the cascade global attention mechanism to obtain Nxm 2 D 3 A dimensional feature matrix;
(4.3) mixing the above-mentioned NXm 2 D 3 Multi-layer perceptron network with dimension feature matrix input based on cascade global attention mechanismA second multi-layer perceptron to obtain NxD H Maintaining a high-level feature matrix;
(5) The above-mentioned NxD H The dimension high-level characteristic matrix adopts maximum pooling to obtain 1 xD H And carrying out object classification on the dimensional global features to obtain a C-dimensional output classification vector C.
Further, as an implementation of the method shown in fig. 1, the embodiment of the invention further provides an object classification device based on the point cloud. The embodiment of the device corresponds to the embodiment of the method, and for convenience of reading, details of the embodiment of the method are not repeated one by one, but it should be clear that the device in the embodiment can correspondingly realize all the details of the embodiment of the method. As shown in fig. 7, the apparatus includes: a first determination unit 21, a second determination unit 22, and a third determination unit 23, wherein
A first determining unit 21 for determining a pair Ji Dian cloud coordinate data of the target object;
a second determining unit 22, configured to determine a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism;
a third determining unit 23, configured to determine global features of the target object based on the high-level feature matrix to determine a classification result.
Illustratively, the above unit is further configured to:
determining point cloud coordinate data based on the target object;
the point cloud coordinate data is spatially transformed based on a spatial transformation network to determine the pair Ji Dian cloud coordinate data.
Illustratively, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
a high-level feature matrix is determined based on the pair Ji Dian cloud coordinate data and the target global feature extraction model based on the global attention mechanism.
Illustratively, the target global feature extraction model based on the global attention mechanism includes: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
By way of example only, and not by way of limitation,
the cascade global attention mechanism is formed by cascading a plurality of global attention mechanisms,
the multi-layer perceptron network is used for extracting the characteristics of the point cloud data,
the feature transformation network is used for aligning the features of the point cloud data.
Illustratively, the determining the high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
performing multi-layer perceptron network processing based on a global attention mechanism on the Ji Dian cloud coordinate data to obtain a low-layer feature matrix;
performing feature transformation network processing based on a cascade global attention mechanism on the low-level feature matrix to acquire an aligned low-level feature matrix;
and performing multi-layer perceptron network processing based on a cascade global attention mechanism on the aligned low-layer feature matrix to acquire a high-layer feature matrix.
Illustratively, the determining the global feature of the target object based on the high-level feature matrix to determine the classification result includes:
carrying out maximum pooling treatment on the high-level feature matrix to obtain global features;
and carrying out full-connection network processing on the global features to classify the target objects.
By means of the technical scheme, the object classification device based on the point cloud is used for solving the problem that classification accuracy and stability are difficult to consider when object classification is carried out based on the point cloud, and the object classification device based on the point cloud is used for determining Ji Dian cloud coordinate data of a target object; determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism; and determining global features of the target object based on the high-level feature matrix to determine a classification result. In the scheme, three key networks of the point cloud classification method based on the global features are redesigned, and in a new network, different global attention mechanisms are organically fused with the original network, so that each point in the point cloud can fully utilize the features of all points in each key stage of feature extraction, the classification precision is improved, and the division of local point cloud or the calculation of the neighborhood of the point is not involved in each stage of feature extraction, so that the classification stability is ensured, and the technical problem that the classification precision and the stability are difficult to consider in the prior art is solved.
The processor includes a kernel, and the kernel fetches the corresponding program unit from the memory. The kernel can be provided with one or more than one, and the object classification method based on the point cloud can be realized by adjusting kernel parameters, so that the problem that classification accuracy and stability are difficult to consider when the object classification is carried out based on the point cloud can be solved.
An embodiment of the present invention provides a computer-readable storage medium including a stored program that, when executed by a processor, implements the above-described point cloud-based object classification method.
The embodiment of the invention provides a processor which is used for running a program, wherein the object classification method based on point cloud is executed when the program runs.
The embodiment of the invention provides electronic equipment, which comprises at least one processor and at least one memory connected with the processor; wherein the processor is configured to call the program instructions in the memory and execute the object classification method based on the point cloud
An embodiment of the present invention provides an electronic device 30, as shown in fig. 8, where the electronic device includes at least one processor 301, and at least one memory 302 and a bus 303 connected to the processor; wherein, the processor 301 and the memory 302 complete communication with each other through the bus 303; the processor 301 is configured to invoke the program instructions in the memory to perform the point cloud based object classification method described above.
The intelligent electronic device herein may be a PC, PAD, cell phone, etc.
The present application also provides a computer program product adapted to perform, when executed on a flow management electronic device, a program initialized with the method steps of:
determining Ji Dian cloud coordinate data of a pair of target objects;
determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism;
and determining global features of the target object based on the high-level feature matrix to determine a classification result.
Further, the method further comprises the following steps:
determining point cloud coordinate data based on the target object;
the point cloud coordinate data is spatially transformed based on a spatial transformation network to determine the pair Ji Dian cloud coordinate data.
Further, the determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
a high-level feature matrix is determined based on the pair Ji Dian cloud coordinate data and the target global feature extraction model based on the global attention mechanism.
Further, the target global feature extraction model based on the global attention mechanism includes: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
Further, the method comprises the steps of,
the cascade global attention mechanism is formed by cascading a plurality of global attention mechanisms,
the multi-layer perceptron network is used for extracting the characteristics of the point cloud data,
the feature transformation network is used for aligning the features of the point cloud data.
Further, the determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism includes:
performing multi-layer perceptron network processing based on a global attention mechanism on the Ji Dian cloud coordinate data to obtain a low-layer feature matrix;
performing feature transformation network processing based on a cascade global attention mechanism on the low-level feature matrix to acquire an aligned low-level feature matrix;
and performing multi-layer perceptron network processing based on a cascade global attention mechanism on the aligned low-layer feature matrix to acquire a high-layer feature matrix.
Further, the determining the global feature of the target object based on the high-level feature matrix to determine a classification result includes:
carrying out maximum pooling treatment on the high-level feature matrix to obtain global features;
and carrying out full-connection network processing on the global features to classify the target objects.
In the foregoing embodiments, the descriptions of the embodiments are focused on, and for those portions of one embodiment that are not described in detail, reference may be made to the related descriptions of other embodiments.
It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
Embodiments of the present application also provide a computer program product comprising computer software instructions which, when run on a processing device, cause the processing device to perform a flow of control of a memory as in the corresponding embodiment of fig. 1.
The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions in accordance with embodiments of the present application are produced in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by a wired (e.g., coaxial cable, fiber optic, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). Computer readable storage media can be any available media that can be stored by a computer or data storage devices such as servers, data centers, etc. that contain an integration of one or more available media. Usable media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid State Disks (SSDs)), among others.
It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.
In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods may be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of elements is merely a logical functional division, and there may be additional divisions of actual implementation, e.g., multiple elements or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.
The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed over a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional unit in each embodiment of the present application may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.
The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be embodied in essence or a part contributing to the prior art or all or part of the technical solution in the form of a software product stored in a storage medium, including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
The above embodiments are only for illustrating the technical solution of the present application, and are not limiting thereof; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.
Claims (8)
1. An object classification method based on point cloud, which is characterized by comprising the following steps:
determining Ji Dian cloud coordinate data of a pair of target objects;
determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism;
determining global features of the target object based on the high-level feature matrix to determine a classification result;
the determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism includes:
determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism-based target global feature extraction model;
The target global feature extraction model based on the global attention mechanism comprises the following steps: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
2. The method as recited in claim 1, further comprising:
determining point cloud coordinate data based on the target object;
the point cloud coordinate data is spatially transformed based on a spatial transformation network to determine the pair Ji Dian cloud coordinate data.
3. The method of claim 1, wherein the step of determining the position of the substrate comprises,
the cascade global attention mechanism is formed by cascading a plurality of global attention mechanisms,
the multi-layer perceptron network is used for extracting the characteristics of the point cloud data,
the feature transformation network is used for aligning the features of the point cloud data.
4. The method of claim 1, wherein the determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism comprises:
performing multi-layer perceptron network processing on the pair Ji Dian cloud coordinate data based on a global attention mechanism to acquire a low-layer feature matrix;
Performing feature transformation network processing based on a cascade global attention mechanism on the low-level feature matrix to acquire an aligned low-level feature matrix;
and performing multi-layer perceptron network processing based on a cascade global attention mechanism on the aligned low-layer feature matrix to acquire a high-layer feature matrix.
5. The method of claim 1, wherein the determining global features of the target object based on the high-level feature matrix to determine classification results comprises:
carrying out maximum pooling treatment on the high-level feature matrix to obtain global features;
and carrying out full-connection network processing on the global features to classify the target objects.
6. An object classification method device based on point cloud is characterized in that,
a first determining unit for determining a pair Ji Dian cloud coordinate data of the target object;
a second determining unit for determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism;
a third determining unit configured to determine global features of the target object based on the high-level feature matrix to determine a classification result;
the determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and a global attention mechanism includes:
Determining a target global feature extraction model based on a global attention mechanism based on a global feature extraction architecture and the global attention mechanism;
determining a high-level feature matrix based on the pair Ji Dian cloud coordinate data and the global attention mechanism-based target global feature extraction model;
the target global feature extraction model based on the global attention mechanism comprises the following steps: a multi-layer perceptron network based on a global attention mechanism, a feature transformation network based on a cascaded global attention mechanism and a multi-layer perceptron network based on a cascaded global attention mechanism.
7. A computer-readable storage medium, characterized in that the computer-readable storage medium comprises a stored program, wherein the steps of the point cloud based object classification method according to any of claims 1 to 5 are implemented when the program is executed by a processor.
8. An electronic device comprising at least one processor and at least one memory coupled to the processor; wherein the processor is adapted to invoke program instructions in the memory to perform the steps of the point cloud based object classification method according to any of the claims 1 to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211076689.7A CN115456064B (en) | 2022-09-05 | 2022-09-05 | Object classification method based on point cloud and related equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211076689.7A CN115456064B (en) | 2022-09-05 | 2022-09-05 | Object classification method based on point cloud and related equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115456064A CN115456064A (en) | 2022-12-09 |
| CN115456064B true CN115456064B (en) | 2024-02-02 |
Family
ID=84300193
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211076689.7A Active CN115456064B (en) | 2022-09-05 | 2022-09-05 | Object classification method based on point cloud and related equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115456064B (en) |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197215A (en) * | 2019-05-22 | 2019-09-03 | 深圳市牧月科技有限公司 | A kind of ground perception point cloud semantic segmentation method of autonomous driving |
| CN111046781A (en) * | 2019-12-09 | 2020-04-21 | 华中科技大学 | Robust three-dimensional target detection method based on ternary attention mechanism |
| CN111242208A (en) * | 2020-01-08 | 2020-06-05 | 深圳大学 | Point cloud classification method, point cloud segmentation method and related equipment |
| CN111931790A (en) * | 2020-08-10 | 2020-11-13 | 武汉慧通智云信息技术有限公司 | Laser point cloud extraction method and device |
| CN112257637A (en) * | 2020-10-30 | 2021-01-22 | 福州大学 | Vehicle-mounted laser point cloud multi-target identification method integrating point cloud and multiple views |
| CN112818999A (en) * | 2021-02-10 | 2021-05-18 | 桂林电子科技大学 | Complex scene 3D point cloud semantic segmentation method based on convolutional neural network |
| KR20210106703A (en) * | 2020-02-21 | 2021-08-31 | 전남대학교산학협력단 | Semantic segmentation system in 3D point cloud and semantic segmentation method in 3D point cloud using the same |
| CN113569979A (en) * | 2021-08-06 | 2021-10-29 | 中国科学院宁波材料技术与工程研究所 | Three-dimensional object point cloud classification method based on attention mechanism |
| US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
| CN113988164A (en) * | 2021-10-21 | 2022-01-28 | 电子科技大学 | Representative point self-attention mechanism-oriented lightweight point cloud target detection method |
| WO2022032823A1 (en) * | 2020-08-10 | 2022-02-17 | 中国科学院深圳先进技术研究院 | Image segmentation method, apparatus and device, and storage medium |
| CN114170465A (en) * | 2021-12-08 | 2022-03-11 | 中国联合网络通信集团有限公司 | Attention mechanism-based 3D point cloud classification method, terminal device and storage medium |
| CN114444613A (en) * | 2022-02-11 | 2022-05-06 | 吉林大学 | Object classification and object segmentation method based on 3D point cloud information |
| CN114445816A (en) * | 2022-01-24 | 2022-05-06 | 内蒙古包钢医院 | Pollen classification method based on two-dimensional image and three-dimensional point cloud |
| CN114913330A (en) * | 2022-07-18 | 2022-08-16 | 中科视语(北京)科技有限公司 | Point cloud component segmentation method and device, electronic equipment and storage medium |
-
2022
- 2022-09-05 CN CN202211076689.7A patent/CN115456064B/en active Active
Patent Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197215A (en) * | 2019-05-22 | 2019-09-03 | 深圳市牧月科技有限公司 | A kind of ground perception point cloud semantic segmentation method of autonomous driving |
| CN111046781A (en) * | 2019-12-09 | 2020-04-21 | 华中科技大学 | Robust three-dimensional target detection method based on ternary attention mechanism |
| CN111242208A (en) * | 2020-01-08 | 2020-06-05 | 深圳大学 | Point cloud classification method, point cloud segmentation method and related equipment |
| KR20210106703A (en) * | 2020-02-21 | 2021-08-31 | 전남대학교산학협력단 | Semantic segmentation system in 3D point cloud and semantic segmentation method in 3D point cloud using the same |
| CN111931790A (en) * | 2020-08-10 | 2020-11-13 | 武汉慧通智云信息技术有限公司 | Laser point cloud extraction method and device |
| WO2022032823A1 (en) * | 2020-08-10 | 2022-02-17 | 中国科学院深圳先进技术研究院 | Image segmentation method, apparatus and device, and storage medium |
| US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
| CN112257637A (en) * | 2020-10-30 | 2021-01-22 | 福州大学 | Vehicle-mounted laser point cloud multi-target identification method integrating point cloud and multiple views |
| CN112818999A (en) * | 2021-02-10 | 2021-05-18 | 桂林电子科技大学 | Complex scene 3D point cloud semantic segmentation method based on convolutional neural network |
| CN113569979A (en) * | 2021-08-06 | 2021-10-29 | 中国科学院宁波材料技术与工程研究所 | Three-dimensional object point cloud classification method based on attention mechanism |
| CN113988164A (en) * | 2021-10-21 | 2022-01-28 | 电子科技大学 | Representative point self-attention mechanism-oriented lightweight point cloud target detection method |
| CN114170465A (en) * | 2021-12-08 | 2022-03-11 | 中国联合网络通信集团有限公司 | Attention mechanism-based 3D point cloud classification method, terminal device and storage medium |
| CN114445816A (en) * | 2022-01-24 | 2022-05-06 | 内蒙古包钢医院 | Pollen classification method based on two-dimensional image and three-dimensional point cloud |
| CN114444613A (en) * | 2022-02-11 | 2022-05-06 | 吉林大学 | Object classification and object segmentation method based on 3D point cloud information |
| CN114913330A (en) * | 2022-07-18 | 2022-08-16 | 中科视语(北京)科技有限公司 | Point cloud component segmentation method and device, electronic equipment and storage medium |
Non-Patent Citations (3)
| Title |
|---|
| GA-NET: Global Attention Network for Point Cloud Semantic Segmentation;Shuang Deng, et al.;arXiv:2107.03101v1 [cs.CV];第1-5页 * |
| Global Attention Mechanism: Retain Information to Enhance Channel-Spatial Interactions;Yichao Liu, et al.;arXiv:2112.05561v1 [cs.CV];第1-6页 * |
| 基于语义信息与动态特征点剔除的SLAM算法;潘海鹏等;浙江理工大学学报（自然科学版）;第47卷(第5期);第764-773页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115456064A (en) | 2022-12-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107945204B (en) | Pixel-level image matting method based on generation countermeasure network | |
| Becker et al. | CNN architecture comparison for radio galaxy classification | |
| Beikmohammadi et al. | SWP-LeafNET: A novel multistage approach for plant leaf identification based on deep CNN | |
| CN111160350A (en) | Portrait segmentation method, model training method, device, medium and electronic equipment | |
| Mohammed Abdelkader et al. | Entropy-based automated method for detection and assessment of spalling severities in reinforced concrete bridges | |
| CN112801047A (en) | Defect detection method and device, electronic equipment and readable storage medium | |
| CN113988164B (en) | Lightweight point cloud target detection method for representative point self-attention mechanism | |
| CN112819008B (en) | Method, device, medium and electronic equipment for optimizing instance detection network | |
| Dengpan et al. | Faster and transferable deep learning steganalysis on GPU | |
| Zhu et al. | Enhanced feature Fusion structure of YOLO v5 for detecting small defects on metal surfaces | |
| Zhang | Application of artificial intelligence recognition technology in digital image processing | |
| Cheng et al. | Water quality monitoring method based on TLD 3D fish tracking and XGBoost | |
| CN111008589B (en) | Face key point detection method, medium, device and computing equipment | |
| JP2023131117A (en) | Joint perception model training, joint perception method, device, and medium | |
| Song et al. | Texture analysis by genetic programming | |
| CN111222558A (en) | Image processing method and storage medium | |
| CN115456064B (en) | Object classification method based on point cloud and related equipment | |
| Belhi et al. | Study and evaluation of pre-trained CNN networks for cultural heritage image classification | |
| Ketsoi et al. | SREFBN: Enhanced feature block network for single‐image super‐resolution | |
| CN113704276A (en) | Map updating method and device, electronic equipment and computer readable storage medium | |
| CN117670860A (en) | Photovoltaic glass defect detection method and device | |
| CN113255512B (en) | Method, apparatus, device and storage medium for living body identification | |
| CN113486718B (en) | Fingertip detection method based on deep multitask learning | |
| CN116958027A (en) | Three-dimensional industrial abnormality detection method and device, storage medium and electronic equipment | |
| Tong et al. | I-YOLO: a novel single-stage framework for small object detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |