Laser radar three-dimensional target detection method based on high-definition map enhancement

Abstract

The invention discloses a laser radar three-dimensional target detection method based on high-definition map enhancement, and the method comprises the following steps: carrying out the voxelization of laser radar data, converting the voxelization into a four-dimensional tensor, and converting the four-dimensional tensor into a needed BEF feature through a sparse encoder; the method comprises the following steps: for high-definition map data, converting the data into required BEF features through a downsampling feature extractor and two shallow feature extractors; and performing data fusion on the BEF features of the laser radar and the BEF features of the high-definition map through an attention module to obtain fused BEF features, and obtaining a final detection result through a decoder. According to the laser radar three-dimensional target detection method based on high-definition map enhancement, the precision and integrity of the point cloud data are improved, the detection effect of the shielded target is optimized, and the real-time processing performance is improved by adopting parallel processing and a lightweight feature extraction network; and a more efficient and accurate solution is provided for laser radar three-dimensional target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser radar three-dimensional target detection, and specifically relates to a laser radar three-dimensional target detection method based on high-definition map enhancement. Background Art

[0002] LiDAR technology emits a laser beam and measures the time difference between it and the object's reflection, thereby accurately calculating the distance between the laser and the target object and building a three-dimensional model. LiDAR can provide high-precision three-dimensional point cloud data in various environments, becoming an important tool in the field of three-dimensional space perception.

[0003] Point cloud data usually contains noise and redundant points, and requires preprocessing steps such as denoising, downsampling, and segmentation. Common denoising methods include statistical filtering, radius filtering, etc. In recent years, deep learning technology has been introduced into the field of point cloud classification and recognition. Point-based neural networks complete the processing of point cloud data by learning global and local geometric features. Although existing technologies have made many advances in point cloud data processing, the sparsity and irregularity of point clouds are still the main challenges in the processing process. How to process massive point cloud data and extract effective geometric and semantic information from it while maintaining data accuracy is still the focus of current research.

[0004] 3D object detection refers to the task of identifying and locating the position, shape and posture of an object in 3D point cloud data. The mainstream object detection algorithms are based on voxel and point. The voxel-based method divides the point cloud data into a fixed-size 3D grid (voxel) and aggregates the point cloud features in each voxel. This method can convert irregular point cloud data into a regular 3D matrix. However, there are two main problems with the voxelization method: First, the selection of voxel resolution is a key issue. Too high a resolution will lead to excessive computation, while too low a resolution will lead to loss of spatial information. Second, quantization errors will be introduced during the voxelization process, especially when processing sparse point clouds, the information loss may be more obvious. Point-based methods: operate directly on the original point cloud without voxelization. By introducing point invariance and capturing local geometric structures, point cloud data can be efficiently processed, especially in sparse point clouds. However, point-based methods have high computational costs, especially when processing large-scale point clouds, and the time and memory consumption are more significant. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a laser radar three-dimensional target detection method based on high-definition map enhancement, which can improve the accuracy and integrity of point cloud data, optimize the detection effect of occluded targets, and improve real-time processing performance.

[0006] The technical solution of the present invention is: a laser radar three-dimensional target detection method based on high-definition map enhancement, comprising the following steps:

[0007] Step 1: voxelize the collected lidar data;

[0008] Step 2: Convert the voxelized lidar data into a four-dimensional tensor;

[0009] Step 3: Input the four-dimensional tensor into the sparse encoder and convert it into the lidar BEF feature. The sparse encoder is a convolutional neural network with a U-Net architecture.

[0010] Step 4: Preprocess the HD map data using GT-Paste;

[0011] Step 5: Input the processed HD map data into the downsampling feature extractor and two shallow feature extractors;

[0012] Step 6: The features extracted by the downsampling feature extractor and the two shallow feature extractors are fused and converted into high-definition map BEF features;

[0013] Step 7: The BEF features of the lidar and the BEF features of the high-definition map are fused through a multimodal fuser and the fused BEF features are output. The multimodal fuser introduces a channel attention processing mechanism and a spatial attention processing mechanism;

[0014] Step 8: Process the fused BEF features through a decoder to output the target detection result. The decoder is a convolutional neural network with a U-Net architecture.

[0015] Furthermore, the detection range of the X axis in the voxelization is 0-70.4m, the detection range of the Y axis is -40-40m, and the detection range of the Z axis is -3-1m; the voxel size along the X, Y and Z axes in the voxelization is 0.05m, 0.05m and 0.1m respectively; the voxel grid resolution in the voxelization is 1408×1600×40; the calculation method of the index value in the voxelization is: let the voxel size be v, the minimum coordinate of the voxel grid be (x min ,y min , z min ), then for a point (x, y, z) in the point cloud, its corresponding voxel index (i, j, k) can be calculated by the following formula:

[0016]

[0017] Among them, [] means rounding down;

[0018] The method of feature aggregation in the voxelization is: taking the maximum value: assuming that there are n points in the voxel V, the feature vector of each point is f i ∈R d , where d is the feature dimension. Then the feature vector f of the voxel V for:

[0019] f V =[max(f i1 ), max(f i2 ), ..., max(f id )],

[0020] Among them, f ij represents the j-th eigenvalue of the ith point.

[0021] Furthermore, the data after voxelization is a three-dimensional tensor, and the expression of the three-dimensional tensor is: (D x , D y , D z ), where D x represents the size of the voxel grid in the X-axis direction, D y Indicates the size of the voxel grid in the Y-axis direction, D z represents the size of the voxel grid in the Z-axis direction; the expression of the four-dimensional tensor formed by adding the reflection intensity to the three-dimensional tensor is: (D x , D y , D z , C), where C represents the reflection intensity; the four-dimensional tensor is standardized after being generated, and the standardization is specifically as follows: for each voxel whose point cloud point number is not T, if its point cloud point number is greater than T, random sampling is performed to make its point cloud T; ​​if its point cloud point number is less than T, the corresponding missing point data is directly set to 0, where T represents the number of target point clouds in each voxel.

[0022] Furthermore, the specific steps of converting the four-dimensional tensor into the laser radar BEF feature through the sparse encoder are as follows: input the four-dimensional tensor (D x , D y , D z , C) processed by multiple sparse convolutional layers; a nonlinear activation function is used after each convolutional layer; the feature map is downsampled through a sparse pooling layer; and merged layer by layer (D z ) dimension, merging the three-dimensional feature information into two-dimensional features; outputting the laser radar BEF features; the expression of the nonlinear activation function is:

[0023] f(x)=max(0,x), where x represents the input value;

[0024] The feature tensor shape of the laser radar BEF feature is (Dx D y , F), where F represents the number of feature channels after fusion.

[0025] Furthermore, the GT-Paste preprocessing method only pastes objects in the target area.

[0026] Furthermore, the shape of the high-definition map data in step 5 is: M×N, where M represents the height of the gridded map and N represents the width of the gridded map; the shape of the high-definition map data output after downsampling in step 5 is: Wherein, M represents the height of the gridded map, N represents the width of the gridded map, S represents the step size of downsampling, and C′ represents the number of channels of the output feature map; the expression of the activation function of the convolutional layer in the downsampling feature extractor is:

[0027] f(x)=max(0,x), where x represents the input value;

[0028] The steps of batch normalization of the convolutional layer in the downsampling feature extractor are:

[0029] Calculate the mean:

[0030]

[0031] Among them, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i represents the value of the i-th sample;

[0032] Calculate the variance:

[0033]

[0034] Among them, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i represents the value of the i-th sample;

[0035] Normalized input:

[0036]

[0037] in, Represents the normalized value, ∈ represents the value is 10 -5 A positive number, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, x i represents the value of the i-th sample;

[0038] Zoom and Pan:

[0039]

[0040] Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, and β represents the translation parameter;

[0041] The expression of batch normalization is:

[0042]

[0043] Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, β represents the translation parameter, σ 2 Represents the variance of the current batch, ∈ represents a value of 10 -5 A positive number.

[0044] Furthermore, the gridding includes: global retrieval and local gridding, and the global retrieval is specifically: first, extracting a local vector map within the detection range from the high-definition map; then, converting the local vector map from the original coordinate system to the local coordinate system of the laser radar; then, defining a rectangular area in the laser radar coordinate system to represent the detection range of the laser radar; then, converting the local vector map to the global coordinate system; then, retrieving and extracting the geographic geometry within the rectangular area in the global map; then, converting the retrieved geographic geometry from the global coordinate system back to the laser radar coordinate system; finally, retrieving the high-definition map data within the detection range again, and aligning its coordinate system with the laser radar; the local gridding is specifically: converting the aligned vector map into a grid map.

[0045] Furthermore, the fusion mode of the feature fusion in step 6 is channel splicing or element-by-element addition, and the high-definition map BEF features after the feature fusion in step 6 are interpolated or deconvolved.

[0046] Furthermore, the architecture of the multimodal fuser includes: data input, channel attention processing, spatial attention processing, fusion and output, the data input includes: input of BEF features of the lidar and BEF features of the high-definition map, the channel attention processing includes the following steps: global average pooling of the lidar BEF features, the pooled features pass through a fully connected layer, the output features of the connection layer pass through the ReLU activation function, the activated output features pass through another fully connected layer, and then pass through the Sigmoid activation function and generate channel attention weights, and the channel attention weights are multiplied by the original lidar BEF features to obtain enhanced features; the spatial attention processing includes the following steps: identity mapping of the high-definition map BEF features, the identity mapped features pass through the first convolutional layer, the output after convolution passes through the ReLU activation function, the activated output passes through the second convolutional layer, and then passes through a convolutional layer to obtain further feature representation; the fusion and output are specifically: the features after channel attention processing and spatial attention processing are fused by element-by-element addition and the results are output.

[0047] Beneficial effects of the present invention: A laser radar three-dimensional target detection method based on high-definition map enhancement of the present invention improves the accuracy and integrity of point cloud data by fusing the acquired laser radar point cloud data with the geometric information in the high-definition map; the entire process includes preprocessing of point cloud data, model construction of convolutional neural network and data enhancement of improved ground truth pasting technology; the detection effect of occluded targets is optimized through multi-perspective data fusion, and parallel processing and lightweight feature extraction networks are used to improve real-time processing performance, and finally the target object category, position coordinates and posture information in the scene are output, providing a more efficient and accurate solution for target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention.

[0049] Figure 1 It is an algorithm flow chart of a laser radar three-dimensional target detection method based on high-definition map enhancement of the present invention.

[0050] Figure 2 It is a schematic diagram of the architecture of a multimodal fuser of a high-definition map-enhanced laser radar three-dimensional target detection method of the present invention.

[0051] Figure 3 A schematic diagram of high-definition map gridding of a laser radar three-dimensional target detection method based on high-definition map enhancement of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the accompanying drawings of 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.

[0053] like Figure 1-2 As shown, a laser radar three-dimensional target detection method based on high-definition map enhancement includes the following steps:

[0054] Step 1: voxelize the collected lidar data;

[0055] Step 2: Convert the voxelized lidar data into a four-dimensional tensor;

[0056] Step 3: Input the four-dimensional tensor into the sparse encoder and convert it into the lidar BEF feature. The sparse encoder is a convolutional neural network with a U-Net architecture.

[0057] Step 4: Preprocess the HD map data using GT-Paste;

[0058] Step 5: Input the processed HD map data into the downsampling feature extractor and two shallow feature extractors;

[0059] Step 6: The features extracted by the downsampling feature extractor and the two shallow feature extractors are fused and converted into high-definition map BEF features;

[0060] Step 7: The BEF features of the lidar and the BEF features of the high-definition map are fused through a multimodal fuser and the fused BEF features are output. The multimodal fuser introduces a channel attention processing mechanism and a spatial attention processing mechanism;

[0061] Step 8: Process the fused BEF features through a decoder to output the target detection result. The decoder is a convolutional neural network with a U-Net architecture.

[0062] Preferably, the detection range of the X axis in the voxelization is 0-70.4m, the detection range of the Y axis is -40-40m, and the detection range of the Z axis is -3-1m; the voxel size along the X, Y and Z axes in the voxelization is 0.05m, 0.05m and 0.1m respectively; the voxel grid resolution in the voxelization is 1408×1600×40; the calculation method of the index value in the voxelization is: let the voxel size be v, the minimum coordinate of the voxel grid be (x min ,y min ,z min), then for a point (x, y, z) in the point cloud, its corresponding voxel index (i, j, k) can be calculated by the following formula:

[0063]

[0064] Among them, [] means rounding down;

[0065] The method of feature aggregation in the voxelization is: taking the maximum value: assuming that there are n points in the voxel V, the feature vector of each point is f i ∈R d , where d is the feature dimension. Then the feature vector f of the voxel V for:

[0066] f V =[max(f i1 ),max(f i2 ),...,max(f id )],

[0067] Among them, f ij represents the j-th eigenvalue of the ith point.

[0068] Preferably, the data after voxel processing is a three-dimensional tensor, and the expression of the three-dimensional tensor is: (D x , D y , D z ), where D x represents the size of the voxel grid in the X-axis direction, D y Indicates the size of the voxel grid in the Y-axis direction, D z represents the size of the voxel grid in the Z-axis direction; the expression of the four-dimensional tensor formed by adding the reflection intensity to the three-dimensional tensor is: (D x , D y , D z , C), where C represents the reflection intensity; the four-dimensional tensor is standardized after being generated, and the standardization is specifically as follows: for each voxel whose point cloud point number is not T, if its point cloud point number is greater than T, random sampling is performed to make its point cloud T; ​​if its point cloud point number is less than T, the corresponding missing point data is directly set to 0, where T represents the number of target point clouds in each voxel.

[0069] Preferably, the specific steps of converting the four-dimensional tensor into the laser radar BEF feature through the sparse encoder are: inputting the four-dimensional tensor (D x , D y , D z , C) processed by multiple sparse convolutional layers; a nonlinear activation function is used after each convolutional layer; the feature map is downsampled through a sparse pooling layer; and merged layer by layer (D z) dimension, merging the three-dimensional feature information into two-dimensional features; outputting the laser radar BEF features; the expression of the nonlinear activation function is:

[0070] f(x)=max(0,x), where x represents the input value;

[0071] The feature tensor shape of the laser radar BEF feature is (D x , D y , F), where F represents the number of feature channels after fusion.

[0072] Preferably, the GT-Paste preprocessing method only pastes objects in the target area.

[0073] Preferably, the shape of the high-definition map data in step 5 is: M×N, where M represents the height of the gridded map and N represents the width of the gridded map; the shape of the high-definition map data output after downsampling in step 5 is: Wherein, M represents the height of the gridded map, N represents the width of the gridded map, S represents the step size of downsampling, and C′ represents the number of channels of the output feature map; the expression of the activation function of the convolutional layer in the downsampling feature extractor is:

[0074] f(x)=max(0,x), where x represents the input value;

[0075] The steps of batch normalization of the convolutional layer in the downsampling feature extractor are:

[0076] Calculate the mean:

[0077]

[0078] Among them, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i represents the value of the i-th sample;

[0079] Calculate the variance:

[0080]

[0081] Among them, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i Represents the value of the i-th sample;

[0082] Normalized input:

[0083]

[0084] in, Represents the normalized value, ∈ represents the value is 10-5 A positive number, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, x i Represents the value of the i-th sample;

[0085] Zoom and Pan:

[0086]

[0087] Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, and β represents the translation parameter;

[0088] The expression of batch normalization is:

[0089]

[0090] Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, β represents the translation parameter, σ 2 Represents the variance of the current batch, ∈ represents a value of 10 -5 A positive number.

[0091] More preferably, Figure 3 As shown, the gridding includes: global retrieval and local gridding, and the global retrieval is specifically: first, extract the local vector map within the detection range from the high-definition map; then, convert the local vector map from the original coordinate system to the local coordinate system of the laser radar; then, define a rectangular area in the laser radar coordinate system to represent the detection range of the laser radar; then, convert the local vector map to the global coordinate system; then, retrieve and extract the geographic geometry within the rectangular area in the global map; then, convert the retrieved geographic geometry from the global coordinate system back to the laser radar coordinate system; finally, retrieve the high-definition map data within the detection range again, and align its coordinate system with the laser radar; the local gridding is specifically: convert the aligned vector map into a grid map.

[0092] Preferably, the fusion mode of the feature fusion in step 6 is channel splicing or element-by-element addition, and the high-definition map BEF features after the feature fusion in step 6 are interpolated or deconvolved.

[0093] Preferably, the architecture of the multimodal fuser includes: data input, channel attention processing, spatial attention processing, fusion and output, the data input includes: input of BEF features of the lidar and BEF features of the high-definition map, the channel attention processing includes the following steps: global average pooling of the lidar BEF features, the pooled features pass through a fully connected layer, the output features of the connection layer pass through the ReLU activation function, the activated output features pass through another fully connected layer, and then pass through the Sigmoid activation function and generate channel attention weights, and the channel attention weights are multiplied by the original lidar BEF features to obtain enhanced features; the spatial attention processing includes the following steps: identity mapping of the high-definition map BEF features, the identity mapped features pass through the first convolutional layer, the output after convolution passes through the ReLU activation function, the activated output passes through the second convolutional layer, and then passes through a convolutional layer to obtain further feature representation; the fusion and output are specifically: the features after channel attention processing and spatial attention processing are fused by element-by-element addition and the results are output.

[0094] Although the specific implementation of the present invention is introduced above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that those non-creative modifications or equivalent replacements of the technical solution of the present invention, which do not depart from the protection scope of the technical solution, should be included in the scope of the claims of the present invention.

Claims

1. A laser radar three-dimensional target detection method based on high-definition map enhancement, characterized in that: The following steps are involved: Step 1: voxelize the collected lidar data; Step 2: Convert the voxelized lidar data into a four-dimensional tensor; Step 3: Input the four-dimensional tensor into the sparse encoder and convert it into the lidar BEF feature. The sparse encoder is a convolutional neural network with a U-Net architecture. Step 4: Preprocess the HD map data using GT-Paste; Step 5: Input the processed HD map data into the downsampling feature extractor and two shallow feature extractors; Step 6: The features extracted by the downsampling feature extractor and the two shallow feature extractors are fused and converted into high-definition map BEF features; Step 7: The BEF features of the lidar and the BEF features of the high-definition map are fused through a multimodal fuser and the fused BEF features are output. The multimodal fuser introduces a channel attention processing mechanism and a spatial attention processing mechanism; Step 8: Process the fused BEF features through a decoder to output the target detection result. The decoder is a convolutional neural network with a U-Net architecture.

2. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1 is characterized in that: The detection range of the X axis in the voxelization is 0-70.4m, the detection range of the Y axis is -40-40m, and the detection range of the Z axis is -3-1m; the voxel size along the X, Y and Z axes in the voxelization is 0.05m, 0.05m and 0.1m respectively; the voxel grid resolution in the voxelization is 1408×1600×40; the calculation method of the index value in the voxelization is: let the voxel size be v, the minimum coordinate of the voxel grid is (x min ,y min ,z min ), then for a point (x, y, z) in the point cloud, its corresponding voxel index (i, j, k) can be calculated by the following formula: Among them, [] means rounding down; The method of feature aggregation in the voxelization is: taking the maximum value: assuming that there are n points in the voxel V, the feature vector of each point is f i ∈R d , where d is the feature dimension. Then the feature vector f of the voxel V for: f V =[max(f i1 ),max(f i2 ),…,max(f id )], Among them, f ij represents the j-th eigenvalue of the ith point.

3. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1 is characterized in that: The data after voxelization is a three-dimensional tensor, and the expression of the three-dimensional tensor is: x , D y , D z ), where D x Indicates the size of the voxel grid in the X-axis direction, D y Indicates the size of the voxel grid in the Y-axis direction, D z represents the size of the voxel grid in the Z-axis direction; the expression of the four-dimensional tensor formed by adding the reflection intensity to the three-dimensional tensor is: (D x , D y , D z , C), where C represents the reflection intensity; the four-dimensional tensor is standardized after being generated, and the standardization is specifically as follows: for each voxel whose point cloud point number is not T, if its point cloud point number is greater than T, random sampling is performed to make its point cloud T; ​​if its point cloud point number is less than T, the corresponding missing point data is directly set to 0, where T represents the number of target point clouds in each voxel.

4. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1, characterized in that: The specific steps of converting the four-dimensional tensor into the laser radar BEF feature through the sparse encoder are: input the four-dimensional tensor (D x , D y , D z , C) processed by multiple sparse convolutional layers; a nonlinear activation function is used after each convolutional layer; The feature map is downsampled through the sparse pooling layer; layer-by-layer merging (D z ) dimension, and fuse the three-dimensional feature information into two-dimensional features; Output the laser radar BEF feature; the expression of the nonlinear activation function is: f(x)=max(0,x), where x represents the input value; The feature tensor shape of the laser radar BEF feature is (D x , D y , F), where F represents the number of feature channels after fusion.

5. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1 is characterized in that: The GT-Paste preprocessing method only pastes objects in the target area.

6. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1 is characterized in that: The shape of the high-definition map data in step 5 is: M×N, where M represents the height of the gridded map and N represents the width of the gridded map; the shape of the high-definition map data output after downsampling in step 5 is: Wherein, M represents the height of the gridded map, N represents the width of the gridded map, S represents the step size of downsampling, and C′ represents the number of channels of the output feature map; the expression of the activation function of the convolutional layer in the downsampling feature extractor is: f(x)=max(0,x), where x represents the input value; The steps of batch normalization of the convolutional layer in the downsampling feature extractor are: Calculate the mean: Among them, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i Represents the value of the i-th sample; Calculate the variance: Among them, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, m represents the number of samples in the current batch, and x i Represents the value of the i-th sample; Normalized input: in, Represents the normalized value, ∈ represents the value is 10 -5 A positive number, σ 2 represents the variance of the current batch, μ represents the mean of the current batch, x i Represents the value of the i-th sample; Zoom and Pan: Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, and β represents the translation parameter; The expression of batch normalization is: Among them, y i represents the value after batch normalization, represents the normalized value, γ represents the scaling parameter, β represents the translation parameter, σ 2 Represents the variance of the current batch, ∈ represents a value of 10 -5 A positive number.

7. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 6 is characterized in that: The gridding includes: global retrieval and local gridding, and the global retrieval is specifically: first, extracting a local vector map within the detection range from the high-definition map; then, converting the local vector map from the original coordinate system to the local coordinate system of the laser radar; then, defining a rectangular area in the laser radar coordinate system to represent the detection range of the laser radar; then, converting the local vector map to the global coordinate system; then, retrieving and extracting the geographic geometry within the rectangular area in the global map; then, converting the retrieved geographic geometry from the global coordinate system back to the laser radar coordinate system; finally, retrieving the high-definition map data within the detection range again, and aligning its coordinate system with the laser radar; the local gridding is specifically: converting the aligned vector map into a grid map.

8. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1 is characterized in that: The fusion mode of the feature fusion in step 6 is channel splicing or element-by-element addition, and the high-definition map BEF features after the feature fusion in step 6 are interpolated or deconvolved.

9. The method for three-dimensional target detection based on laser radar and high-definition map enhancement according to claim 1, characterized in that: The architecture of the multimodal fuser includes: data input, channel attention processing, spatial attention processing, fusion and output, the data input includes: input of BEF features of the laser radar and BEF features of the high-definition map, the channel attention processing includes the following steps: global average pooling of the laser radar BEF features, the pooled features pass through a fully connected layer, the output features of the connection layer pass through the ReLU activation function, the activated output features pass through another fully connected layer, and then pass through the Sigmoid activation function and generate channel attention weights, and the channel attention weights are multiplied by the original laser radar BEF features to obtain enhanced features; the spatial attention processing includes the following steps: identity mapping of the high-definition map BEF features, the identity mapped features pass through the first convolution layer, the output after convolution passes through the ReLU activation function, the activated output passes through the second convolution layer, and then passes through a convolution layer to obtain further feature representation; the fusion and output are specifically: the features after channel attention processing and spatial attention processing are fused by element-by-element addition and the results are output.

Images