Three-dimensional point cloud target detection method based on automatic driving environment

Abstract

The invention discloses a three-dimensional point cloud target detection method based on an automatic driving environment. The method comprises the following implementation steps: firstly, acquiring original point clouds of vehicles, pedestrians, riders and the like in an automatic driving scene from a KITTI data set, carrying out data preprocessing, and dividing point cloud data into vertical point columns (Pillars); the method comprises the following steps of: firstly, carrying out point cloud clustering on a Pill in a feature extraction module by using a point column clustering sampling method designed by the invention, clustering point clouds in the Pill into N clusters by using DBSCAN density clustering, carrying out center sampling on each cluster, and processing each Pill by using a mixed pooling Dill Convolution Layer (MPDC) Layer. And then, in the backbone network, in combination with a space channel synergy enhancement module (SCS-EM), the context sensing capability of the network is enhanced. And finally, training the network, calculating a loss function to optimize a detection model, and performing final three-dimensional target detection by using an SSD detection head. Through experimental verification, the method effectively improves the accuracy of three-dimensional point cloud target detection.

Description

Technical Field

[0001] The present invention relates to the field of image target detection, and in particular to a three-dimensional point cloud target detection method based on an autonomous driving environment. Background Art

[0002] With the rapid development of autonomous driving technology, modern urban traffic environment has put forward higher requirements on the environmental perception ability of autonomous vehicles. Autonomous vehicles need to make accurate decisions and plans by perceiving the surrounding environment to ensure the safety and reliability of driving. Three-dimensional point cloud target detection technology can obtain the three-dimensional coordinate information of objects in the environment through sensors such as lidar, and realize accurate perception and positioning of targets. Among them, PointPillars is a powerful voxel-based three-dimensional point cloud target detection framework, which can quickly process a large amount of point cloud data without significantly sacrificing accuracy, which is crucial for real-time decision-making of autonomous vehicles in complex environments. Compared with traditional two-dimensional image technology, three-dimensional point cloud target detection technology can provide more comprehensive and accurate environmental perception information, thereby providing a more reliable basis for decision-making and control of autonomous vehicles.

[0003] Existing point cloud detection models are often affected by noise and density differences when processing sparse and unevenly distributed point cloud data, resulting in inaccurate feature extraction and limitations in extracting rich feature representations. When using random discarding methods and traditional pooling methods (such as maximum pooling), important information may be lost. Therefore, the performance of target detection is slightly insufficient in understanding complex scenes and multi-scale targets. For example, patent CN202211496628-Three-dimensional point cloud target detection method, device and storage medium, the clustering algorithm (such as spectral clustering) used in this method may be less efficient when processing large-scale data, and requires a preset number of clusters, which may not be suitable for dynamic scenes. And the performance and results of the clustering algorithm are highly dependent on the parameter settings. For example, spectral clustering requires the selection of appropriate similarity metrics and clustering neighbor parameters. Improper parameter settings may lead to unsatisfactory clustering effects or even complete failure. Patent CN202211104980-A 3D point cloud target detection method based on a second-order multi-attention mechanism of a point column uses a second-order multi-attention mechanism (including a second-order point attention mechanism, a second-order channel attention mechanism, and a pseudo-image space attention mechanism) to increase the complexity of the model. Although a new attention mechanism is proposed, there is still a high dependence on the quality of the input features in the feature extraction process. If the input features are inaccurate, the accuracy of the detection results may be affected. Patent CN202410379713-A point cloud target detection method based on an attention mechanism and a feature enhancement structure uses a variety of attention mechanisms and feature enhancement structures, such as point attention, channel attention, and spatial attention, which increases the complexity of the model. This not only increases the computational overhead, but also increases the difficulty of implementation and debugging. In addition, this method depends on the quality and density of the point cloud data. If the input data is sparse or noisy, it may affect the effect of the attention mechanism, thereby affecting the detection accuracy. Summary of the invention

[0004] In view of the above analysis, the present invention proposes a three-dimensional point cloud target detection method based on an autonomous driving environment by improving the PointPillars algorithm, aiming to solve the limitations such as large scale differences in feature representation, complex feature laws, and loss of spatial information.

[0005] To achieve the above object, the technical solution of the present invention is as follows: A three-dimensional point cloud target detection method based on an autonomous driving environment comprises the following steps:

[0006] Step 1: Get the original point cloud data for data preprocessing and divide it into Pillar units: Divide the input point cloud into multiple Pillar units. Each Pillar is a three-dimensional small unit cell obtained by evenly dividing the point cloud with a certain step size on the XY plane.

[0007] Step 2: In the feature extraction module, perform feature extraction on the point cloud data:

[0008] First, the point cloud data in multiple Pillar units are processed. For Pillars with less than N point clouds, the zero-filling method is used to fill them to N points. For Pillars with more than N points, the point column clustering sampling method is used to select N representative points.

[0009] Then use the mixed pooling and dilated convolution layer to extract features from each Pillar;

[0010] Finally, the extracted features are converted into a two-dimensional pseudo image according to Pillar's arrangement.

[0011] Step 3: The pseudo image is passed into the backbone network:

[0012] First, three downsamplings are performed. At the final stage of the three downsamplings, the spatial channel collaborative enhancement module is introduced to obtain the feature map Y c , the spatial channel collaborative enhancement module includes combining a spatial attention module and a channel attention module;

[0013] Then upsampling and feature concatenation are performed to convert the feature map Y c Restore it to the same dimension size as the input and perform feature concatenation to obtain the enhanced feature map.

[0014] Step 4: Train the neural network on the enhanced feature map, calculate the loss function to optimize the detection model, and use the SSD (Single Shot Multi Box Detector) detection head for the final 3D target detection.

[0015] Furthermore, in the above step 2, the method of point column clustering sampling is: clustering the point cloud in Pillar into N clusters using DBSCAN density clustering, and then performing center sampling on each cluster to select representative points.

[0016] Furthermore, in the above step 2, the construction method of the hybrid pooling extended convolutional layer is:

[0017] First, perform maximum pooling and average pooling to extract different features of the input feature map X:

[0018] X max =MaxPool(X) (1)

[0019] X avg =AvgPool(X) (2)

[0020] Among them, MaxPool is the maximum pooling operation, and AvgPool is the average pooling operation.

[0021] For the pooled feature map Xmax and X avg Apply dilated convolution:

[0022] X' max =DilatedConv(X max ,r) (3)

[0023] X' avg =DilatedConv(X avg ,r) (4)

[0024] Among them, DilatedConv represents the dilated convolution operation, and r represents the dilation rate.

[0025] Finally, the two pooled feature maps are fused through the results of the extended convolution to capture richer contextual information:

[0026] X'=α·X' max +(1-α)·X' avg (5)

[0027] Among them, α is a learnable parameter or preset ratio, which is used to control the fusion ratio of the maximum pooling features and the average pooling features. The final output feature map X' will contain the features extracted by the maximum pooling and average pooling, as well as the spatial context information enhanced by the dilated convolution.

[0028] Furthermore, in the above step 3, the construction method of the spatial channel collaborative enhancement module is as follows:

[0029] First, the module takes the input feature map X c Perform horizontal and vertical global mean pooling to capture global spatial information.

[0030] Global mean pooling in horizontal direction:

[0031]

[0032] Global mean pooling in vertical direction:

[0033]

[0034] Among them, h and w represent the position of the feature map in the horizontal and vertical directions respectively.

[0035] Then, the module generates an attention map by fusing the global information in these two directions and using convolution and activation functions. The fused feature f is processed by a shared network F1:

[0036]

[0037] in, represents the concatenation of horizontal and vertical global mean pooling features, δ is the activation function, and F1 is the combination of 1×1 convolution and subsequent activation function. Along the spatial dimension, f is split into f h ∈R C / r×H×1 and f w ∈R C / r×1×W , then convolution is performed to increase the dimension, and then combined with the sigmoid activation function to obtain the final attention vector g h ∈R C ×H×1 and g w ∈R C×1×W .

[0038] g h =σ(F h (f h )) (9)

[0039] g w =σ(F w (f w )) (10)

[0040] Where σ is the sigmoid activation function, F h and F w is a 1×1 convolution operation applied in different directions, f h and f w are the horizontal and vertical components of f. The module takes the input feature map X c Using the channel attention mechanism, through global average pooling and maximum pooling operations, and then through a multilayer perceptron (MLP) network. The output of the channel attention branch can be expressed as:

[0041]

[0042] Among them, σ is the sigmoid activation function, AvgPool and MaxPool represent global average pooling and maximum pooling respectively, and W0 and W1 are the weight matrices in MLP.

[0043] At the final output, SCS-EM combines the spatial attention map and the channel attention map to the input feature map X c Weighted, and add the identity matrix I to maintain the original strength of the feature to obtain the final output feature map Y c :

[0044] Y c (i,j)=X c (i,j)·g h (i) g w (j)+X c(i,j)·M c (12)

[0045] Here, i and j represent the spatial positions of the feature maps, respectively, and · represents element-wise multiplication.

[0046] Furthermore, in the above step 2, the value range of N is [20,100].

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) In order to solve the problem of information loss when extracting features from sparse point cloud data, the present invention designs a point column clustering sampling method that combines DBSCAN density clustering and center sampling methods to process a large amount of sparse and non-uniformly distributed point cloud data (greater than N), cleverly avoiding the problem of key information loss that may be caused by traditional random sampling. The original PointPillars network encodes the point cloud in the Pillar and directly enters the network without clustering, center sampling, etc. The improved network of the present invention better retains point cloud information and can more effectively process sparse and non-uniform point cloud data. The features calculated based on the core points of the cluster are usually more distinguishable than edge points, ensuring that each body column contains the most representative points, which can effectively reduce information loss.

[0049] (2) To address the problem of large scale differences when extracting point cloud features, the present invention proposes a mixed pooling dilated convolution layer (MPDC Layer), which combines maximum pooling, average pooling and applies a new feature extraction technology of dilated convolution. It increases feature diversity while retaining spatial information, thereby improving the recognition ability and robustness of the model in complex environments.

[0050] (3) To address the problem of complex feature patterns and single feature expressions, the present invention designs a Space-Channel Synergistic Enhancement Module (SCS-EM), which not only enhances the model's focus on important areas, but also maintains a comprehensive understanding of the overall features, making the module perform better in image processing tasks. This module can further enhance the spatial and channel information of the feature map, making the network capture key information more accurately, thereby improving the accuracy and efficiency of target detection.

[0051] (4) The above three improvements work together to improve the performance of the PointPillars network in processing point cloud data. By better preserving point cloud information, enhancing feature diversity, and improving feature expression capabilities, the improved network performs well in processing sparse and non-uniformly distributed point cloud data, target recognition in complex environments, and feature expression and capture, significantly improving the accuracy and efficiency of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is the overall flow chart of the network model of the present invention;

[0053] Figure 2 It is a network model structure diagram of the present invention;

[0054] Figure 3 This is a schematic diagram of the DBSCAN clustering algorithm;

[0055] Figure 4 Schematic diagram of the method for extending the convolutional layer for mixed pooling;

[0056] Figure 5 The figure is a comparison diagram of the detection results of the present invention and the PointPillars model in one embodiment. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and main points of the present invention clearer, the present invention will be further described in detail below in conjunction with the attached examples. Obviously, the described embodiments are only part of the embodiments of the present invention, which are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0058] like Figure 1 Shown is the overall flow chart of the present invention. The basic idea of ​​the present invention is to first select the original point cloud data from the KITTI dataset, and then divide the input point cloud into multiple small three-dimensional cells on the XY plane, and then divide it into vertical point columns (Pillars) along the Z axis, and use zero padding or DBSCAN clustering to deal with the unevenness of the number of point clouds in the Pillars. Then, a mixed pooling and extended convolution layer is used to process the point cloud of each Pillar, and these features are converted into a two-dimensional pseudo image for convolutional network processing. Afterwards, in the backbone network, the feature expression is enhanced by combining spatial and channel attention through the spatial channel collaborative enhancement module, and then the size of the feature map is restored by upsampling and feature splicing to obtain the enhanced feature map. Finally, the enhanced feature map is trained by the neural network, and the SSD detection head is used to detect the three-dimensional target to obtain the final detection result. The overall network model structure diagram of this process is shown in the figure. Figure 2 shown.

[0059] Based on the above basic ideas, the present invention provides a three-dimensional point cloud target detection method based on an improved PointPillars algorithm network. Figure 2 The specific implementation steps are as follows.

[0060] Step 1: Select the original point cloud data from the autonomous driving scene dataset, preprocess it and divide it into Pillar units:

[0061] First, the input point cloud data is normalized and the coordinates of the point cloud data are converted into a fixed range.

[0062] Pillar unit division is to divide the input point cloud into multiple Pillar units. Each Pillar is a three-dimensional small unit cell obtained by evenly dividing the point cloud with a certain step size on the XY plane.

[0063] In this embodiment, the point cloud data range is set to [0, -39.68, -3, 69.12, 39.68, 1], and the voxel size is set to [0.16, 0.16, 4].

[0064] Step 2: Perform feature extraction on the point cloud data in the feature extraction module, and convert the extracted features into a two-dimensional pseudo image according to the Pillar arrangement so that the subsequent convolutional network can process it more conveniently.

[0065] Step 1.1: First, process the point cloud data in multiple Pillar units:

[0066] (1) For Pillars with less than N point clouds, the zero-filling method is used to fill the points to N points. The max_points threshold set by the present invention is 35, that is, N=35;

[0067] (2) For Pillars with more than N points, the point-column clustering sampling method is used to select N representative points:

[0068] The point column clustering sampling method designed by the present invention uses the DBSCAN density clustering method and center sampling to retain key features. The specific operation process is as follows:

[0069] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a typical density-based spatial clustering algorithm. The algorithm diagram is as follows: Figure 3As shown in the figure, a schematic diagram of noise points, core points, and boundary points is marked. First, two parameters are set, the radius ε and the minimum number of points MinPts; then, for each point, the number of adjacent points in its ε neighborhood is calculated; if a point has at least MinPts points in its ε neighborhood, it is marked as a core point; all points that are not core points but are in the ε neighborhood of the core point are marked as boundary points; the remaining points are marked as noise points. The algorithm then defines the area formed by each core point or a group of boundary points connected to the core point as a cluster, and the core points can be connected through boundary points. After the cluster is established, the noise point is not included in any cluster, and the algorithm continues to iterate until all points are visited and marked, so the DBSCAN algorithm can effectively cluster according to the local density characteristics of the data.

[0070] The present invention first adopts the DBSCAN clustering method to identify all dense areas of sample points, and regards these dense areas as independent clusters. For each body column, if the number of points therein exceeds the set max_points threshold, the network will perform DBSCAN clustering. The max_points threshold set by the present invention is 35. In this process, the present invention only retains points marked as core samples. This operation is intended to reduce information loss and maintain the key features of the data. The original PointPillars network encodes the point cloud in the Pillar and directly enters the network without clustering, center sampling and other processing. The improved network of the present invention better retains point cloud information and can more effectively process sparse and uneven point cloud data. And the features calculated based on the core points of the cluster are usually more discriminative than edge points.

[0071] The specific process is to first perform density clustering processing, and then select the center point through center sampling. The point column clustering sampling method proposed in the present invention effectively identifies and retains dense areas in point cloud data, which often contain rich and important information.

[0072] Step 2.2: In order to further enhance the network's ability to express features and enhance the network's adaptability to different scales and density changes in point cloud data, this paper designs a mixed pooling dilated convolution layer (MPDC Layer). Figure 4 As shown in the figure, the overall processing flow is described, which extracts features from the point cloud and then uses them for target detection.

[0073] First, perform maximum pooling and average pooling to extract different features of the input feature map X:

[0074] X max =MaxPool(X) (1)

[0075] Xavg =AvgPool(X) (2)

[0076] Among them, MaxPool is the maximum pooling operation, and AvgPool is the average pooling operation.

[0077] For the pooled feature map X max and X avg Apply dilated convolution:

[0078] X' max =DilatedConv(X max ,r) (3)

[0079] X' avg =DilatedConv(X avg ,r) (4)

[0080] Among them, DilatedConv represents the dilated convolution operation, and r represents the dilation rate.

[0081] Finally, the two pooled feature maps are fused through the results of the extended convolution to capture richer contextual information:

[0082] X'=α·X' max +(1-α)·X' avg (5)

[0083] Among them, α is a learnable parameter or preset ratio, which is used to control the fusion ratio of the maximum pooling features and the average pooling features. The final output feature map X' will contain the features extracted by the maximum pooling and average pooling as well as the spatial context information enhanced by the dilated convolution.

[0084] Step 2.3: Finally, the feature vectors of all body columns are arranged into a pseudo-image according to their positions in space. Each pixel of this pseudo-image corresponds to a feature vector of a body column and serves as the input of the 2D convolutional neural network.

[0085] Step 3: The pseudo image is sent to the backbone network, where the feature expression is enhanced and the final 3D target detection is performed. The specific steps include:

[0086] Step 3.1: In the final stage of three downsampling in the backbone network, the Space-Channel Synergistic Enhancement Module (SCS-EM) is introduced, which further enhances the feature expression of the network by combining spatial attention and channel attention.

[0087] The spatial channel enhancement module proposed in this paper aims to optimize the ability of 3D convolutional neural networks to process complex spatial data. Horizontal and vertical global mean pooling is designed in the SCS-EM module, combined with an additional channel attention branch to highlight the key features of the image in multiple dimensions. In this way, the module can capture the global spatial distribution of the image in detail, greatly improving the accuracy of feature extraction.

[0088] First, the module takes the input feature map X c Perform horizontal and vertical global mean pooling to capture global spatial information.

[0089] Global mean pooling in horizontal direction:

[0090]

[0091] Global mean pooling in vertical direction:

[0092]

[0093] Among them, h and w represent the position of the feature map in the horizontal and vertical directions respectively.

[0094] Then, the module generates an attention map by fusing the global information of these two directions and using 1x1 convolution and activation function. The fused feature f is processed by a shared network F1:

[0095]

[0096] in, represents the concatenation of horizontal and vertical global mean pooling features, δ is the activation function, and F1 is the combination of 1×1 convolution and subsequent activation function. Along the spatial dimension, f is split into f h ∈R C / r×H×1 and f w ∈R C / r×1×W , then convolution is performed to increase the dimension, and then combined with the sigmoid activation function to obtain the final attention vector g h ∈R C ×H×1 and g w ∈R C×1×W .

[0097] g h =σ(F h (f h )) (9)

[0098] g w =σ(F w (f w )) (10)

[0099] Where σ is the sigmoid activation function, F h and F w is a 1×1 convolution operation applied in different directions, f h and f w are the horizontal and vertical components of f.

[0100] At the same time, the other branch of the module inputs the feature map X c Using the channel attention mechanism, through global average pooling and maximum pooling operations, and then through a multilayer perceptron (MLP) network. The output of the channel attention branch can be expressed as:

[0101]

[0102] Among them, σ is the sigmoid activation function, AvgPool and MaxPool represent global average pooling and maximum pooling respectively, and W0 and W1 are the weight matrices in MLP.

[0103] At the final output, SCS-EM combines the spatial attention map and the channel attention map to the input feature map X c Weighted, and add the identity matrix I to maintain the original strength of the feature to obtain the final output feature map Y c :

[0104] Y c (i,j)=X c (i,j)·g h (i) g w (j)+X c (i,j)·M c (12)

[0105] Among them, i and j represent the spatial positions of the feature map, and · represents element-by-element multiplication. This method not only enhances the model's attention to important areas, but also maintains a comprehensive understanding of the overall features, making the module perform better in image processing tasks.

[0106] Step 3.2: Preliminary processed feature map Y c The feature is upsampled using transposed 2D convolution, which aims to restore the low-dimensional feature map to the same dimensional size as the input point cloud data. After upsampling, feature maps of different scales are spliced ​​using feature fusion technology to combine low-level features with high-level features, enhance the ability of feature expression, and obtain enhanced feature maps. This process provides the detection network with rich semantic information and accurate spatial positioning capabilities.

[0107] Step 4: The enhanced feature map is trained through a neural network, and the loss function is calculated to optimize the detection model. Using the detection head with the SSD (Single Shot Multi Box Detector) architecture, the target detection network can directly predict the position, category and other attributes of the target on the feature map, thereby achieving accurate 3D target detection and obtaining the final detection result.

[0108] The final visual detection result of the present invention is as follows Figure 5 As shown, the figure shows the detection results in two different autonomous driving scenarios, scenario one is an urban scene, and scenario two is a highway scene. The bright red box represents the 3D detection box of the car (Car), the purple box represents the 3D detection box of the pedestrian (Pedestrian), and the dark red box represents the 3D detection box of the cyclist (Cyclist). In the experiment, the point cloud in the 3D detection box in the point cloud scene is dyed with the corresponding color to facilitate the observation of the visualization effect. In addition, the missed detection and false detection are marked with a yellow border. It can be seen from the detection results of scene one that the present invention can accurately identify small targets and occluded vehicles in the distance compared with the detection results of the PointPillars algorithm. It can be seen from the detection results of scene two that the PointPillars algorithm will miss detections for people and cars in the distance, and the PointPillars algorithm will miss detections for vehicles that do not exist on the right. Therefore, the missed detection and false detection of the present invention are significantly less than the PointPillars algorithm.

[0109] Table 1 shows the comparison of the performance indicators of the present invention and other algorithms, in which the results with the highest detection accuracy are bolded. The mean average precision (mAP) of the improved PointPillars algorithm in the Car category on the KITTI test set is almost higher than that of other algorithms, indicating that the algorithm proposed in the present invention is advanced.

[0110] Table 1 Comparison of mAP of different algorithms in Car category

[0111]

[0112]

[0113] The above is an explanation of the specific implementation of the present invention, rather than a limitation of the present invention. Those skilled in the relevant technical field can also make various equivalent technical solutions without departing from the scope of the present invention, so all equivalent technical solutions should be included in the patent protection scope of the present invention.

Claims

1. A three-dimensional point cloud target detection method based on an autonomous driving environment, characterized in that: The following steps are involved: Step 1: Get the original point cloud data, perform data preprocessing and divide it into Pillar units: divide the input point cloud into multiple Pillar units, each Pillar is a three-dimensional small unit cell obtained by evenly dividing the point cloud with a certain step length on the XY plane; Step 2: In the feature extraction module, perform feature extraction on the point cloud data: First, the point cloud data in multiple Pillar units are processed. For Pillars with less than N point clouds, the zero-filling method is used to fill them to N points. For Pillars with more than N points, the point column clustering sampling method is used to select N representative points. Then use the mixed pooling and dilated convolution layer to extract features from each Pillar; Finally, the extracted features are converted into a two-dimensional pseudo image according to Pillar’s ​​arrangement; Step 3: The pseudo image is passed into the backbone network: First, three downsamplings are performed. At the final stage of the three downsamplings, the spatial channel collaborative enhancement module is introduced to obtain the feature map Y c , the spatial channel collaborative enhancement module includes a combination of a spatial attention module and a channel attention module; Then upsampling and feature concatenation are performed to convert the feature map Y c Restore to the same dimension size as the input and perform feature concatenation to obtain an enhanced feature map; Step 4: Train the neural network on the enhanced feature map, calculate the loss function to optimize the detection model, and use the SSD (Single Shot MultiBox Detector) detection head for the final 3D target detection.

2. The three-dimensional point cloud target detection method based on the autonomous driving environment according to claim 1, characterized in that: In the step 2, the point column clustering sampling method is to cluster the point cloud in Pillar into N clusters using DBSCAN density clustering, and then perform center sampling on each cluster to select representative points.

3. The three-dimensional point cloud target detection method based on an autonomous driving environment according to claim 1, characterized in that: In the step 2, the construction method of the mixed pooling extended convolutional layer is: First, perform maximum pooling and average pooling to extract different features of the input feature map X: X max =MaxPool(X) (1) X avg =AvgPool(X) (2) Among them, MaxPool is the maximum pooling operation, and AvgPool is the average pooling operation; For the pooled feature map X max and X avg Apply dilated convolution: X' max =DilatedConv(X max ,r) (3) X' avg =DilatedConv(X avg ,r) (4) Among them, DilatedConv represents the dilated convolution operation, and r represents the dilation rate; Finally, the two pooled feature maps are fused through the results of the extended convolution to capture richer contextual information: X'=α·X' max +(1-a)·X' avg (5) Among them, α is a learnable parameter or preset ratio, which is used to control the fusion ratio of maximum pooling features and average pooling features; the final output feature map X' will contain the features extracted by maximum pooling and average pooling, as well as the spatial context information enhanced by the dilated convolution.

4. A three-dimensional point cloud target detection method based on an autonomous driving environment according to claim 2 or 3, characterized in that: In step 3, the construction method of the spatial channel collaborative enhancement module is as follows: First, the module takes the input feature map X c Perform horizontal and vertical global mean pooling to capture global spatial information Global mean pooling in horizontal direction: Global mean pooling in vertical direction: Among them, h and w represent the position of the feature map in the horizontal and vertical directions respectively; Then, the module generates an attention map by fusing the global information in these two directions and using convolution and activation functions. The fused feature f is processed by a shared network F1: in, represents the concatenation of horizontal and vertical global mean pooling features, δ is the activation function, and F1 is the combination of 1×1 convolution and subsequent activation function. Along the spatial dimension, f is split into f h ∈R C / r×H×1 and f w ∈R C / r×1×W , then convolution is performed to increase the dimension, and then combined with the sigmoid activation function to obtain the final attention vector g h ∈R C ×H×1 and g w ∈R C×1×W . g h =σ(F h (f h )) (9) g w =σ(F w (f w )) (10) Where σ is the sigmoid activation function, F h and F w is a 1×1 convolution operation applied in different directions, f h and f w are the horizontal and vertical components of f. The module takes the input feature map X c Using the channel attention mechanism, through global average pooling and maximum pooling operations, and then through a multilayer perceptron (MLP) network. The output of the channel attention branch can be expressed as: Among them, σ is the sigmoid activation function, AvgPool and MaxPool represent global average pooling and maximum pooling respectively, and W0 and W1 are the weight matrices in MLP. At the final output, SCS-EM combines the spatial attention map and the channel attention map to the input feature map X c Weighted, and add the identity matrix I to maintain the original strength of the feature to obtain the final output feature map Y c : Y c (i,j)=X c (i,j)·g h (i)·g w (j)+X c (i,j)·M c (12) Here, i and j represent the spatial positions of the feature maps, respectively, and · represents element-wise multiplication.

5. The three-dimensional point cloud target detection method based on the autonomous driving environment according to claim 4, characterized in that: In the step 2, the value range of N is [20, 100].

Images