3D point cloud target detection method based on graph convolution
By using the AlexNet algorithm to perform graph convolution in 3D point cloud object detection, the problem of unclear results classification in the existing technology is solved, and the detection accuracy and calculation efficiency are improved.
Patent Information
- Application Number
- CN202411691902.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-03-18
AI Technical Summary
The existing 3D point cloud object detection method does not provide a clear result classification method, resulting in low detection accuracy and large calculation amount.
The 3D point cloud object detection method based on graph convolution is used to classify the detection results through the AlexNet algorithm, and the convolution layer, pooling layer, fully connected layer, Dropout and Softmax activation functions are used to extract and classify the features.
A clearer result classification is achieved, detection accuracy is improved, computational amount is reduced, and the spatial structure of point cloud information is retained.
Smart Images
Figure CN119648990A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and in particular to a 3D point cloud target detection method based on graph convolution. Background Art
[0002] 3D point cloud refers to a collection of a large number of three-dimensional points. Each point has X, Y and Z coordinates (sometimes RGB information and laser reflection intensity information, etc., required for different scenarios), which respectively represent its position in three-dimensional space. These points are usually obtained from the real world through laser scanning technology (such as lidar) or structured light scanning technology (such as Kinect or other depth sensors). 3D point clouds are usually used to represent and capture objects, scenes or environments in a three-dimensional environment. They can be static, such as a three-dimensional model of a building or a three-dimensional reconstruction of a landscape, or dynamic, such as a point cloud of a moving object captured over time. Existing 3D object detection methods can be divided into two types, one is based on multimodal fusion, and the other is based on pure point cloud. The method based on multimodal fusion first maps the point cloud to different perspectives, and then fuses the representations under different modes. However, this mapping will lose a lot of spatial structure information, resulting in low accuracy and more computation. Therefore, most detection frameworks currently use the second method based on pure point clouds. This framework can be divided into two categories. One is the point-based method, which takes the original point cloud as input and uses iterative clustering operations to learn the representation of the original point cloud. However, its frequent clustering operations bring more computational complexity, and the uneven density of the point cloud also leads to a decrease in its performance. The other type is a voxel-based method, which divides the point cloud into 3D voxel space and uses stacked 3D convolutional layers to learn spatial features. After searching, it is found that the typical 3D point cloud target detection methods in the prior art are as follows: A 3D point cloud target detection method based on graph convolution, the method comprising: inputting and preprocessing the original point cloud data; voxelizing the preprocessed point cloud data and inputting it into a trained 3D point cloud target detection model for detection and analysis; performing statistical analysis and post-processing on the detection results; the 3D point cloud target detection model comprises a 3D sparse convolutional neural network, a proposal generation network, a global dynamic graph convolution module, a local static graph convolution module and a 3D detection head module; its main feature is to use graph convolution to establish the relationship between 3D targets, extract richer dependency features and retain the original spatial information of the target, so that the extracted features can better express the point cloud information.
[0003] In summary, the existing 3D point cloud target detection methods do not provide a clear result classification method. To address the above problems, the existing methods need to be improved. Summary of the invention
[0004] The purpose of the present invention is to provide a 3D point cloud target detection method based on graph convolution to solve the problem that the existing 3D point cloud target detection method proposed in the above background technology does not provide a clear result classification method.
[0005] To achieve the above object, the present invention provides the following technical solution: a 3D point cloud target detection method based on graph convolution, comprising the following steps:
[0006] S1, cloud data acquisition stage: access the Internet to obtain 3D point cloud data in real time;
[0007] S2, detection model construction stage: construct a 3D point cloud data detection model for detection based on the existing original data pairs;
[0008] S3, data import model stage: input the real-time cloud data obtained in stage S1 into the data detection model constructed in stage S2, so as to obtain the target data detection result;
[0009] S4, detection result classification stage: use the classification algorithm to classify the target data detection results and store them.
[0010] Preferably, the construction process of the S2 detection model construction phase includes data division, voxel features, compact processing, feature extraction, information input and information calculation.
[0011] Preferably, the data division in the S2 detection model building stage is performed by importing existing original point cloud data and classifying the original point cloud data into a training set and a test set.
[0012] Preferably, the voxel feature in the S2 detection model construction stage is expressed as voxelizing the point cloud data in the training set converted from the Yuanshu data, and inputting the voxelized point cloud data into the sparse convolutional network for high-dimensional sparse feature extraction to obtain high-dimensional sparse voxel features of different scales. At the same time, sparse convolution uses GEMM to accelerate matrix operations. In spconv, the torch.mm_out function is directly called, which relies on the GEMM that torch has already done. GEMM does not change the amount of calculation, but is achieved by writing the loop separately and making more use of cache storage instead of memory access.
[0013] Preferably, the compact processing in the S2 detection model building stage is performed by processing voxel features and obtaining a processing solution therefrom.
[0014] Preferably, the feature extraction in the S2 detection model construction stage is manifested as processing the processing scheme, and the processing form includes screening and filtering of data characteristics, and extracting the features of the constructed graph to obtain semantic information.
[0015] Preferably, the information input in the S2 detection model construction stage is performed by inputting the obtained semantic information into the classification branch and the regression branch to obtain category information.
[0016] Preferably, the information calculation in the S2 detection model construction stage is manifested as the calculation of the information completion loss function, and parameter adjustment is performed until the loss function reaches a minimum value, that is, the model training is completed.
[0017] Preferably, the classification in the S4 stage is completed by the AlexNet algorithm, and the AlexNet algorithm includes a convolutional layer, a pooling layer, a fully connected layer, Dropout and a Softmax activation function. The convolutional layer in the AlexNet algorithm is responsible for feature extraction, the pooling layer is used for dimensionality reduction, the fully connected layer is used for classification, Dropout is used to prevent overfitting, and Softmax is used to output probability distribution.
[0018] Preferably, the main idea of the Softmax regression algorithm is: for a training set consisting of m labeled samples {(x (1) ,y (1) )...(x (m) ,y (m) )}, where X is the sample feature, Y is the category corresponding to each sample, and the class label y (i) ∈{1,2,...,k}, use the hypothesis function to estimate the probability value of the test sample X to which it belongs, P(y=jx′) takes the category with the largest probability as the predicted category of the test sample, and the softmax regression algorithm is implemented on the Spark platform as follows: first, broadcast the model parameters obtained in the training phase to each computing node on the master node; then read the data set to be classified from HDFS to generate an RDD whose internal data structure is the customer information model, and specify the number of partitions of the RDD according to the size of the data set; then parse the data through the map operation to generate an RDD with a custom data structure inside. Because no iteration is required, there is no need to call the cache to persist the data set at this time; then calculate the probability of each data belonging to each category on each slave node according to the model parameters, and take the category with the largest probability as the predicted category of the data item; finally, output the classification result.
[0019] Compared with the prior art, the beneficial effects of the present invention are: the 3D point cloud target detection method based on graph convolution,
[0020] The present invention solves the problem that the existing 3D point cloud target detection method does not provide a clear result classification method through the AlexNet algorithm. First, the features of the image are extracted through a convolution layer, each convolution layer is composed of multiple convolution kernels and an activation function, the convolution kernel scans the input image, and extracts features by calculating the convolution operation, and the activation function introduces nonlinearity to increase the expression ability of the network; then the dimension and calculation amount of the feature map are reduced through a pooling layer; the commonly used pooling operation is maximum pooling, that is, the maximum value is selected in a fixed window as the value after pooling; the extracted features are classified through a fully connected layer, the fully connected layer expands the feature map into a one-dimensional vector, and maps it to the output category through linear transformation and activation function; then some neurons are randomly inactivated in the fully connected layer, so as to reduce the overfitting phenomenon and improve the generalization ability of the model; finally, the output of the network is converted into a probability distribution through a Softmax activation function, thereby realizing the classification of the result image. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 This is a schematic diagram of the overall process of a 3D point cloud target detection method based on graph convolution in the present invention;
[0022] Figure 2 This is a content composition diagram of the S2 stage of a 3D point cloud target detection method based on graph convolution of the present invention;
[0023] Figure 3 This is a schematic diagram of the AlexNet network structure of a 3D point cloud target detection method based on graph convolution in the present invention. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0025] See also Figure 1-3 , the present invention provides a technical solution:
[0026] Embodiment 1
[0027] A 3D point cloud target detection method based on graph convolution includes the following steps:
[0028] S1, cloud data acquisition stage: access the Internet to obtain 3D point cloud data in real time;
[0029] S2, detection model construction stage: construct a 3D point cloud data detection model for detection based on the existing original data pairs;
[0030] S3, data import model stage: input the real-time cloud data obtained in stage S1 into the data detection model constructed in stage S2, so as to obtain the target data detection result;
[0031] S4, detection result classification stage: use the classification algorithm to classify the target data detection results and store them.
[0032] In a further example, the construction process of the S2 detection model construction phase includes data partitioning, voxel features, compact processing, feature extraction, information input and information calculation.
[0033] In a further example, data division in the S2 detection model building stage is performed by importing existing raw point cloud data and classifying the raw point cloud data into a training set and a test set.
[0034] In a further example, the voxel feature in the S2 detection model construction stage is expressed as voxelizing the point cloud data in the training set converted from the Yuanshu data, and inputting the voxelized point cloud data into the sparse convolutional network for high-dimensional sparse feature extraction to obtain high-dimensional sparse voxel features of different scales. At the same time, sparse convolution uses GEMM to accelerate matrix operations. In spconv, directly calling the torch.mm_out function means relying on the GEMM that torch has already done. GEMM does not change the amount of calculation, but is achieved by writing the loop separately and making more use of cache storage instead of memory access.
[0035] In a further example, the compact processing in the S2 detection model building stage is performed by processing voxel features and obtaining a processing solution therefrom.
[0036] In a further example, feature extraction in the S2 detection model construction stage is manifested as processing the processing solution, and the processing form includes screening and filtering of data characteristics, and extracting the features of the constructed graph to obtain semantic information.
[0037] In a further example, the information input in the S2 detection model construction phase is performed by inputting the obtained semantic information into the classification branch and the regression branch to obtain category information.
[0038] In a further example, the information calculation in the S2 detection model construction stage is manifested as the calculation of the information loss function, and the parameters are adjusted until the loss function reaches the minimum value, that is, the model training is completed.
[0039] In a further example, the classification in the S4 stage is completed by the AlexNet algorithm, and the AlexNet algorithm includes a convolutional layer, a pooling layer, a fully connected layer, a Dropout and a Softmax activation function. The convolutional layer in the AlexNet algorithm is responsible for feature extraction, the pooling layer is used for dimensionality reduction, the fully connected layer is used for classification, the Dropout is used to prevent overfitting, and the Softmax is used to output probability distribution.
[0040] AlexNet algorithm provides eight layers of convolution. Figure 3 As shown in the figure, in the first convolution layer, 48 11×11×3 convolution kernels are used to calculate 48 55×55 feature maps, and another 48 11×11×3 convolution kernels are used to calculate another 48 55×55 feature maps. The convolution steps of these two branches are both 4, and the image size is reduced from 224′224 to 55×55 through convolution. After the first convolution layer, local response normalization (LRN) and maximum pooling with a step size of 2 and a pooling window of 3×3 are performed, and the feature map size of the pooled output is 27×27;
[0041] In the second convolution layer, two groups of 128 5×5×48 convolution kernels are used to convolve the two sets of input feature maps, and two groups of 128 27×27 feature maps are output. After the second convolution layer, local response normalization and maximum pooling with a step size of 2 and a pooling window of 3×3 are performed, and the size of the feature map output by the pooling is 13×13;
[0042] The third convolution layer combines the two sets of feature maps into one. Use 192 3×3×256 convolution kernels to perform convolution operations on all input feature maps, and then use another 192 3×3×256 convolution kernels to perform convolution operations on all input feature maps, outputting two sets of 192 13×13 feature maps each;
[0043] In the fourth convolution layer, 192 3×3×192 convolution kernels are used to perform convolution operations on the two sets of input feature maps respectively;
[0044] In the fifth convolution layer, 128 3×3×192 convolution kernels are used to perform convolution operations on the two sets of input feature maps. After the fifth convolution layer, a maximum pooling with a stride of 2 and a pooling window of 3×3 is performed, and the size of the feature map output by the pooling is 6×6;
[0045] The sixth and seventh fully connected layers have two groups of neurons (2048 neurons each);
[0046] The eighth fully connected layer outputs 1000 features and sends them to softmax, which outputs the probability of classification.
[0047] In a further example, the main idea of the Softmax regression algorithm is: for a training set consisting of m labeled samples {(x (1) ,y (1) )...(x (m) ,y (m) )}, where X is the sample feature, Y is the category corresponding to each sample, and the class label y (i) ∈{1,2,...,k}, use the hypothesis function to estimate the probability value of the test sample X to which it belongs, P(y=jx′) takes the category with the largest probability as the predicted category of the test sample, and the softmax regression algorithm is implemented on the Spark platform as follows: first, broadcast the model parameters obtained in the training phase to each computing node on the master node; then read the data set to be classified from HDFS to generate an RDD whose internal data structure is the customer information model, and specify the number of partitions of the RDD according to the size of the data set; then parse the data through the map operation to generate an RDD with a custom data structure inside. Because no iteration is required, there is no need to call the cache to persist the data set at this time; then calculate the probability of each data belonging to each category on each slave node according to the model parameters, and take the category with the largest probability as the predicted category of the data item; finally, output the classification result.
[0048] Embodiment 2
[0049] This embodiment is a further description of the above embodiment. It should be understood that this embodiment includes all the aforementioned technical features and is further described in detail.
[0050] The image classification method in the S4 detection result classification stage also includes the SIFT algorithm, which extracts feature vectors from the image and then inputs these feature vectors as data into the classifier to complete the classification of the image. The SIFT algorithm principle includes scale space extreme value detection, key point location, feature direction determination, descriptor generation and feature matching.
[0051] The SIFT algorithm flow includes the following steps:
[0052] First, prepare the input data: expression count matrix and cell attribute signal; secondly, calculate the cell-to-cell similarity matrix: select the appropriate kernel function and calculate the cell-to-cell similarity based on signal mapping; then calculate the projection matrix of the expression matrix on the cell-to-cell similarity through matrix multiplication, and subtract the projection matrix from the original data to obtain the filtered data; after filtering, the hidden structure is exposed to facilitate subsequent analysis.
[0053] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A 3D point cloud target detection method based on graph convolution, characterized in that: The following steps are involved: S1, cloud data acquisition stage: access the Internet to obtain 3D point cloud data in real time; S2, detection model construction stage: construct a 3D point cloud data detection model for detection based on the existing original data pairs; S3, data import model stage: input the real-time cloud data obtained in stage S1 into the data detection model constructed in stage S2, so as to obtain the target data detection result; S4, detection result classification stage: use the classification algorithm to classify the target data detection results and store them.
2. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The construction process of the S2 detection model construction phase includes data division, voxel features, compact processing, feature extraction, information input and information calculation.
3. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The data division is performed by importing existing original point cloud data and classifying the original point cloud data into a training set and a test set.
4. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: In the S2 detection model construction stage, the voxel feature is expressed by voxelizing the point cloud data in the training set converted from the Yuanshu data, and inputting the voxelized point cloud data into the sparse convolutional network for high-dimensional sparse feature extraction to obtain high-dimensional sparse voxel features of different scales. At the same time, sparse convolution uses GEMM to accelerate matrix operations. In spconv, directly calling the torch.mm_out function means relying on the GEMM that torch has already done. GEMM does not change the amount of calculation, but is achieved by writing the loop separately and making more use of cache storage instead of memory access.
5. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The compact processing in the S2 detection model building stage is performed by processing voxel features and obtaining a processing solution based on them.
6. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The feature extraction in the S2 detection model construction stage is manifested as processing the processing solution, and the processing form includes screening and filtering of data characteristics, and extracting the features of the constructed graph to obtain semantic information.
7. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The information input in the S2 detection model construction stage is performed by inputting the obtained semantic information into the classification branch and the regression branch to obtain category information.
8. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The information calculation in the S2 detection model construction stage is manifested as the calculation of the information loss function, and parameter adjustment is performed until the loss function reaches the minimum value, that is, the model training is completed.
9. The 3D point cloud target detection method based on graph convolution according to claim 1, characterized in that: The classification in the S4 stage is completed by the AlexNet algorithm, and the AlexNet algorithm includes a convolutional layer, a pooling layer, a fully connected layer, Dropout and a Softmax activation function. The convolutional layer in the AlexNet algorithm is responsible for feature extraction, the pooling layer is used for dimensionality reduction, the fully connected layer is used for classification, Dropout is used to prevent overfitting, and Softmax is used to output probability distribution.
10. The 3D point cloud target detection method based on graph convolution according to claim 9, characterized in that: The main idea of the Softmax regression algorithm is: for a training set consisting of m labeled samples {(x (1) ,y (1) )...(x (m) ,y (m) )}, where X is the sample feature, Y is the category corresponding to each sample, and the class label y (i) ∈{1,2,...,k}, use the hypothesis function to estimate the probability value of the test sample X to which it belongs, P(y=jx′) takes the category with the largest probability as the predicted category of the test sample, and the softmax regression algorithm is implemented on the Spark platform as follows: first, broadcast the model parameters obtained in the training phase to each computing node on the master node; then read the data set to be classified from HDFS to generate an RDD whose internal data structure is the customer information model, and specify the number of partitions of the RDD according to the size of the data set; then parse the data through the map operation to generate an RDD with a custom data structure inside. Because no iteration is required, there is no need to call the cache to persist the data set at this time; then calculate the probability of each data belonging to each category on each slave node according to the model parameters, and take the category with the largest probability as the predicted category of the data item; finally, output the classification result.