A point cloud 3D target detection method based on an Inception module and an ASPP module

Abstract

The application belongs to the technical field of point cloud 3D target detection, and in particular to a point cloud 3D target detection method based on an Inception module and an ASPP module, comprising the following steps: step 1: point cloud data preprocessing, cropping point clouds, extracting target point clouds in a region of interest, enhancing point cloud data using a data enhancement technique, and obtaining high-quality point cloud data after preprocessing. The application integrates a 3D sparse Inception module in a 3D sparse backbone network and introduces an ASPP module in a 2D dense backbone network, significantly improving the network's ability to extract multi-scale features; the 3D sparse Inception module is good at capturing information from multi-scale spatial features, while the ASPP module expands the network's receptive field by processing different sizes of convolution kernels in parallel, enhancing the adaptability to multi-scale targets, which is particularly effective for improving the detection accuracy of small targets such as pedestrians, while maintaining high detection performance for large targets, enhancing the flexibility and robustness of the model.

Description

Technical Field

[0001] This invention relates to the field of point cloud 3D target detection technology, specifically a point cloud 3D target detection method based on the Inception module and the ASPP module. Background Technology

[0002] With the development of autonomous driving technology, the demand for precise environmental perception is increasing. 3D object detection technology can provide richer information than traditional 2D object detection; it can not only identify what an object is, but also determine its specific position and orientation in three-dimensional space. This is crucial for ensuring the safety of autonomous vehicles, as they must understand their surroundings in real time to make rapid and accurate decisions.

[0003] To improve detection accuracy, voxel-based 3D target detection methods divide point cloud data into a three-dimensional mesh (voxels), with each voxel summarizing point cloud information within its spatial region. This method utilizes 3D sparse convolutional networks to extract features, maps these features onto a bird's-eye view (BEV) pseudo-image, and finally applies 2D detection techniques to identify the target.

[0004] However, current voxel-based detection methods typically use fixed-size convolutional kernels to extract features. This design not only limits the model's ability to detect objects at different scales but also hinders the effective integration of long-range feature information. Especially when dealing with small targets such as bicycles and pedestrians with sparse point cloud data distribution, the target detection accuracy is low. Therefore, we propose a point cloud 3D target detection method based on the Inception module and the ASPP module. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a point cloud 3D target detection method based on the Inception module and the ASPP module, which solves the problems mentioned in the background.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0009] A point cloud 3D object detection method based on the Inception module and ASPP module includes the following steps:

[0010] Step 1: Point cloud data preprocessing, point cloud cropping, extraction of target point cloud within the region of interest, and data augmentation techniques to enhance the point cloud data to obtain high-quality preprocessed point cloud data;

[0011] Step 2: Encode the voxel features of the preprocessed point cloud data to generate a 3D sparse feature map of the voxels;

[0012] Step 3: Use a 3D sparse convolutional neural network as the backbone network, insert a 3D sparse Inception module into the network, fuse multi-scale features, and output a 3D sparse tensor.

[0013] Step 4: Project the 3D sparse tensor onto a 2D plane to generate Bird's Eye View (BEV) features;

[0014] Step 5: Use a 2D convolutional neural network containing ASPP modules as the backbone network to further learn BEV features and output a 2D dense feature tensor.

[0015] Step 6: Input the 2D dense feature tensor into the object detection network to obtain the object detection results.

[0016] Furthermore, step 1 preprocesses the 3D point cloud data by cropping the original point cloud data according to the range (Z,Y,X) to obtain a foreground point cloud containing the target. The cropped foreground point cloud is then enhanced using affine transformation and GT-sampling (ground real-time sampling) data augmentation techniques.

[0017] Furthermore, the voxel feature encoding step in step 2 includes:

[0018] Step 2.1: Spatial partitioning, setting the voxel size to (v D ,v H ,v W If the entire point cloud space is divided into a regular voxel grid, then the number of grids in the depth direction of the point cloud is D = Z / v. D The number of grid cells in the height direction of the point cloud is H = Y / v H The number of grids in the width direction of the point cloud is W = X / v W Voxels containing point clouds are denoted as non-empty voxels, and voxels not containing point clouds are denoted as empty voxels.

[0019] Step 2.2: Voxel encoding. The maximum number of points per voxel is set to N. For non-empty voxels with more than N points, N points are retained through random sampling. For voxels with fewer than N points, zero vectors are used to fill in the N points. The local mean of the point cloud within the voxel is calculated as the voxel's centroid. The point cloud features within the voxel are encoded, including the point coordinates, reflection intensity, and the relative position of the point cloud and the voxel's centroid. The length of the encoded feature is denoted as L. The feature tensor of each voxel can be denoted as (N, L). The feature P of each encoded point is shown in the formula:

[0020]

[0021] Step 2.3: Feature extraction. A fully connected network (FCN) layer is used to map the point features to a new feature space. A max-pooling layer is used to aggregate the point cloud features within a voxel, setting the number of feature channels for each voxel to C, and generating a voxel feature vector. The point cloud features can be represented by a (D, H, W, C) dimensional tensor, and the voxel feature V can be represented as:

[0022] P′=FCN(P)

[0023] V = MaxPooling(P′).

[0024] Furthermore, the 3D sparse backbone network in step 3 is as follows:

[0025] By transforming the feature channel number to C1 using two 3D sparse residual blocks, we obtain a (D,H,W,C1)-dimensional tensor denoted as F1; then, we downsample F1 using two 3D sparse residual blocks to obtain... The tensor is denoted as F2; ​​F2 is downsampled and multi-scale features are fused using two 3D sparse Inception modules to obtain... The tensor is denoted as F3; F3 is downsampled and multi-scale features are fused using two 3D sparse Inception modules to obtain... The tensor is denoted as F4.

[0026] Furthermore, step 4 projects the 3D sparse tensor of the point cloud into two-dimensional bird's-eye view features. The specific steps are as follows:

[0027] First, transpose the depth feature vector of the input 3D feature tensor. Then, perform a matrix multiplication between the transposed result and the channel feature vector to obtain a result of size [missing information]. tensor; will The tensor splicing length is The channel vector; replace the original channel vector to generate The BEV tensor.

[0028] Furthermore, step 5 uses a 2D convolutional neural network containing ASPP modules as the backbone network to extract deeper BEV features, specifically including:

[0029] For the input The 3D BEV tensor is transformed using one FCN layer to increase the number of feature channels. The transformed number of feature channels is denoted as C2. The tensor is then used; next, three residual blocks are used to extract channel features, resulting in... 1D tensor; using a hollow spatial pyramid (ASPP) with 4 parallel branches to fuse multi-scale features; generating BEV tensor.

[0030] Furthermore, in step 6, the detection network used includes three detection heads, each responsible for a different prediction task: a classification prediction head for identifying the target's category; a bounding box prediction head for predicting the target's bounding box; and an orientation angle prediction head for estimating the target's orientation angle. The bounding box prediction loss is denoted as L. reg The classification prediction loss is denoted as L. cls The direction angle prediction loss is denoted as L. dir The total loss is denoted as L. total ;

[0031] For the detection box prediction loss L reg Parametrically convert the 3D real bounding box to (x g ,y g ,z g ,l g ,h g ,w g ,θ g The predicted anchor box is parameterized as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of the target regression 7 denoted as (x t ,y t ,z t ,l t ,h t ,w t ,θ t ),

[0032]

[0033] Where x, y, and z are the center coordinates; w, l, and h are the width, length, and height, respectively; θ is the yaw rotation about the z-axis; and the subscripts t, a, and g represent the coded value, anchor value, and actual value, respectively; and It is the diagonal line at the bottom of the anchor box, α represents the weight parameter, i represents the target index, and L is the detection box prediction loss. Reg It can be represented as:

[0034]

[0035] Classification loss L cls Expressed as a formula:

[0036] L cls =-α t (1-p t ) γ log(p t )

[0037] In the formula, p t It is the estimated probability of the model, and α and γ are the parameters of the focus loss;

[0038] The direction angle prediction loss is denoted as L. dir Expressed as a formula:

[0039] L dir =SmoothL1(sin(θ) p -θ t ))

[0040] In the formula, the subscript p represents the predicted value;

[0041] β1, β2, and β3 represent the weight parameters, and the total loss L total Recorded as:

[0042] L total =β1L cls +β2L reg +β3L dir .

[0043] Furthermore, the 3D sparse Inception module used in the 3D sparse backbone network in step 3 specifically involves: extracting receptive field features of different sizes through four parallel 3D submanifold convolutional branches, with sizes of 1×1×1, 3×3×3, 5×5×5, and 7×7×7, respectively; wherein the receptive fields of 5×5×5 and 7×7×7 are achieved by combining 5×1×5 with 1×5×1 and 7×1×7 with 1×7×1 submanifold sparse convolutional kernels, respectively; concatenating the features of these four branches along the channel dimension to make the number of channels four times that of the original input; and using a 1×1×1 convolutional kernel to perform multi-scale information fusion and restore the number of channels to the original input.

[0044] Furthermore, the ASPP module in step 5 has the following main structure: It includes a 1×1 convolutional kernel and three 3×3 convolutional kernels with different dilation rates (1, 6, and 12 respectively), which operate in parallel. The feature maps generated by the four parallel convolutional branches are concatenated along the channel dimension. The number of channels in the output feature map is four times the number of channels in the input feature map.

[0045] (III) Beneficial Effects

[0046] Compared with existing technologies, this invention provides a point cloud 3D object detection method based on the Inception module and the ASPP module, which has the following beneficial effects:

[0047] This invention integrates a 3D sparse Inception module into a 3D sparse backbone network and introduces an ASPP module into a 2D dense backbone network, significantly improving the network's ability to extract multi-scale features. This innovation achieves efficient detection of targets of different sizes with almost no increase in computational or parameter burden. The 3D sparse Inception module excels at capturing information from multi-scale spatial features, while the ASPP module expands the network's receptive domain by processing convolutional kernels of different sizes in parallel, enhancing its adaptability to multi-scale targets. This is particularly effective in improving the detection accuracy of small targets such as pedestrians, while maintaining high detection performance for large targets, thus enhancing the model's flexibility and robustness. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0049] Figure 2 This is a diagram showing the overall network structure of a point cloud 3D target detection method based on the Inception module and the ASPP module of the present invention.

[0050] Figure 3 This is a structural diagram of the 3D sparse Inception module of the method of the present invention;

[0051] Figure 4 This is a schematic diagram of the ASPP module structure in the method of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example

[0054] This invention underwent experimental evaluation. The experiment split 3712 samples from 7481 labeled samples in the KITTI dataset as the training set and 3769 samples as the test set. The overall flowchart of this invention is as follows: Figure 1 As shown, the network structure schematic diagram of the present invention is as follows: Figure 2 As shown, an embodiment of the present invention proposes a point cloud 3D target detection method based on the Inception module and the ASPP module, which includes the following steps:

[0055] Step 1: Data preprocessing. The original point cloud data is cropped along the z×y×x axis in the range of [-3,1]×[-40,40]×[0,70.4]m, retaining only the foreground point cloud data related to the target. At this time, Z=4m, Y=80m, X=70.4m. Affine transformation and GT-sampling (ground real-time sampling) data augmentation techniques are used to enhance the cropped point cloud.

[0056] Step 2: Encode the voxel features of the preprocessed point cloud data to generate a 3D sparse feature map of the voxels. The specific steps are as follows:

[0057] Step 2.1: Spatial partitioning. Set the voxel size to (0.1m, 0.05m, 0.05m) and divide the point cloud space into a regular voxel grid. The number of grids D in the depth direction is 40, the number of grids H in the height direction is 1600, and the number of grids W in the width direction is 1408. Voxels containing point clouds are denoted as non-empty voxels, and voxels not containing point clouds are denoted as empty voxels.

[0058] Step 2.2: Voxel encoding. Set the maximum number of points N for each voxel to 5. For voxels with more than 5 points, randomly sample 5 points. For non-empty voxels with fewer than 5 points, fill them with zero vectors to 5 points. Calculate the local mean of the point cloud coordinates as the centroid of the voxel. Encode the point cloud features within the voxel, including the point coordinates, reflection intensity, and the relative position of the point and the voxel centroid. The encoding length is 7. Each voxel can be represented by a (5, 7) tensor. The feature P of each point after encoding is shown in the formula:

[0059]

[0060]

[0061] Step 2.3: Feature extraction. A fully connected network (FCN) layer is used to map the point features P to a new feature space, and the number of output feature channels C is denoted as 16. A max-pooling layer is used to aggregate the point cloud features within voxels, generating a voxel feature vector of length 16. The (D,H,W,C) dimensional point cloud features can be represented by a (40,1600,1408,16) dimensional tensor, and the voxel features V can be expressed by the formula:

[0062] P′=FCN(P)

[0063] V = MaxPooling(P′).

[0064] Step 3: Use a 3D sparse convolutional neural network as the backbone network, insert a 3D sparse Inception module into the network, fuse multi-scale features, and output a 3D sparse tensor, such as... Figure 2 The 3D sparse backbone network shown is as follows:

[0065] By transforming the number of input feature channels using two 3D sparse residual blocks, the number of channels after transformation is C1 = 16, resulting in a (40, 1600, 1408, 16) dimensional tensor, denoted as F1. F1 is downsampled using two 3D sparse residual blocks, resulting in a (20, 800, 704, 32) dimensional tensor, denoted as F2. F2 is downsampled and multi-scale features are fused using two 3D sparse Inception modules, resulting in a (10, 400, 352, 64) dimensional tensor, denoted as F3. F3 is downsampled and multi-scale features are fused using two 3D sparse Inception modules, resulting in a (5, 200, 176, 64) dimensional tensor, denoted as F4.

[0066] Step 4: Project the 3D sparse tensor onto a 2D plane to generate Bird's Eye View (BEV) features, specifically:

[0067] Transpose the depth feature vector of the input 3D feature tensor F4, multiply the transposed result with the channel feature vector to obtain a tensor of size (5,64); concatenate the (5,64) tensor to obtain a channel vector of length 320; replace the channel vector of F4 to generate a (200,176,320) dimensional BEV tensor.

[0068] Step 5: Using a 2D convolutional neural network containing ASPP modules as the backbone network, further learn BEV features and output a 2D dense feature tensor, such as... Figure 2 The 2D backbone network shown is as follows:

[0069] For the input (200, 176, 320) dimensional BEV features, use one FCN layer to transform the number of feature channels. The transformed number of feature channels C2 is denoted as 128, and the output is a (200, 176, 128) dimensional tensor F6. Process F6 with three residual blocks to obtain a (200, 176, 128) dimensional tensor F7. Use a hollow spatial pyramid (ASPP) with four parallel branches to fuse multi-scale features, and generate a (200, 176, 512) dimensional BEV tensor F8.

[0070] Step 6: Input the 2D dense feature tensor F8 into the object detection network to obtain the object detection results; specifically, use AnchorHead as the detection network, which contains three detection heads, each responsible for a different prediction task: a classification prediction head for identifying the object's category; a bounding box prediction head for predicting the object's bounding box; and an orientation angle prediction head for estimating the object's orientation angle; the classification prediction loss is denoted as L. cls The detection box prediction loss is denoted as L. reg The direction angle prediction head is denoted as L. dir The total loss is denoted as L. total ;

[0071] For the detection box prediction loss L reg Parametrically convert the 3D real bounding box to (x g ,y g ,z g ,l g ,h g ,w g ,θ g The predicted anchor box is parameterized as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of the target regression 7 denoted as (x t ,y t ,z t ,l t ,h t ,w t ,θ t ),

[0072]

[0073] θ t =θ g -θ a

[0074] Where x, y, and z are the center coordinates; w, l, and h are the width, length, and height, respectively; θ is the yaw rotation about the z-axis; and the subscripts t, a, and g represent the coded value, anchor value, and actual value, respectively; and It is the diagonal line at the bottom of the anchor box, α represents the weight parameter, i represents the target index, and L is the detection box prediction loss. Reg Represented as:

[0075]

[0076] Classification loss L cls Expressed as a formula:

[0077] L cls =-α t (1-p t ) γ log(p t )

[0078] In the formula, p t This is the estimated probability of the model, and α and γ are the parameters of the focus loss. During training, α = 0.25 and γ = 2 are used.

[0079] The direction angle prediction loss is denoted as L.dir Expressed as a formula:

[0080] L dir =SmoothL1(sin(θ) p -θ t ))

[0081] In the formula, the subscript p represents the predicted value;

[0082] Total loss L total Recorded as:

[0083] L total =β1L cls +β2L reg +β3L dir

[0084] β1, β2, and β3 represent weight parameters. During training, β1 is 1.0, β2 is 2.0, and β3 is 0.2.

[0085] like Figure 3 As shown in the embodiment, the main structure of the 3D sparse Inception module is as follows:

[0086] Receptive field features of different sizes are extracted through four parallel 3D submanifold convolutional branches: 1×1×1, 3×3×3, 5×5×5, and 7×7×7. The 5×5×5 and 7×7×7 receptive fields are achieved by combining 5×1×5 with 1×5×1 and 7×1×7 with 1×7×1 submanifold sparse convolutional kernels, respectively. The features from these four branches are concatenated along the channel dimension, quadrupling the number of channels in the original input. Finally, a 1×1×1 convolutional kernel is used for multi-scale information fusion to restore the original input's channel count.

[0087] like Figure 4 As shown in the embodiment, the main structure of the ASPP module is as follows:

[0088] The ASPP module consists of a 1×1 convolutional kernel and three 3×3 convolutional kernels with different dilation rates (1, 6, and 12), which operate in parallel. The feature maps generated by the four parallel convolutional branches are concatenated along the channel dimension. The number of channels in the output feature map is four times the number of channels in the input feature map.

[0089] A point cloud detection model was trained, and the model's 3D detection results were output. 3D average precision (R40) was chosen as the evaluation metric, and the evaluated targets included Car, Cyclist, and Pedestrian. Easy indicates that the minimum bounding box height of the evaluated target is 40 pixels, with no occlusion and a maximum truncation of 15%; Moderate indicates that the minimum bounding box height of the evaluated target is 25 pixels, with partial occlusion and a maximum truncation of 30%; Hard indicates that the minimum bounding box height of the evaluated target is 25 pixels, with a maximum occlusion level of barely visible and a maximum truncation of 50%. The detection results are shown in Table 1.

[0090] Table 1. Detection results of the KITTI dataset.

[0091]

[0092] The results show that the method described in this application exhibits superior performance across all difficulty levels and detection categories. In particular, for pedestrian and cyclist detection, compared to other models such as PointPillars, Voxel-NeXt, and SECOND, our model achieves high average accuracy of 62.56% and 64.17% respectively under medium difficulty conditions, demonstrating effective detection capability for these two types of targets.

[0093] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A point cloud 3D target detection method based on the Inception module and the ASPP module, characterized in that, Includes the following steps: Step 1: Point cloud data preprocessing, point cloud cropping, extraction of target point cloud within the region of interest, and data augmentation techniques to enhance the point cloud data to obtain high-quality preprocessed point cloud data; Step 2: Encode the voxel features of the preprocessed point cloud data to generate a 3D sparse feature map of the voxels; Step 3: Use a 3D sparse convolutional neural network as the backbone network, insert a 3D sparse Inception module into the network, fuse multi-scale features, and output a 3D sparse tensor. Step 4: Project the 3D sparse tensor onto a 2D plane to generate BEV features; Step 5: Use a 2D convolutional neural network containing ASPP modules as the backbone network to further learn BEV features and output a 2D dense feature tensor. Step 6: Input the 2D dense feature tensor into the object detection network to obtain the object detection results.

2. The point cloud 3D target detection method based on the Inception module and ASPP module according to claim 1, characterized in that: Step 1 preprocesses the 3D point cloud data by cropping the original point cloud data according to the range (Z,Y,X) to obtain a foreground point cloud containing the target. The cropped foreground point cloud is then enhanced using affine transformation and GT-sampling (ground real-time sampling) data augmentation techniques.

3. The point cloud 3D target detection method based on the Inception module and ASPP module according to claim 2, characterized in that: The voxel feature encoding step in step 2 includes: Step 2.1: Spatial partitioning, setting the voxel size to (v D ,v H ,v W If the entire point cloud space is divided into a regular voxel grid, then the number of grids in the depth direction of the point cloud is D = Z / v. D The number of grid cells in the height direction of the point cloud is H = Y / v H The number of grids in the width direction of the point cloud is W = X / v W Voxels containing point clouds are denoted as non-empty voxels, and voxels not containing point clouds are denoted as empty voxels. Step 2.2: Voxel encoding. The maximum number of points per voxel is set to N. For non-empty voxels with more than N points, N points are retained through random sampling. For voxels with fewer than N points, zero vectors are used to fill in the N points. The local mean of the point cloud within the voxel is calculated as the voxel's centroid. The point cloud features within the voxel are encoded, including the point coordinates, reflection intensity, and the relative position of the point cloud and the voxel's centroid. The length of the encoded feature is denoted as L. The feature tensor of each voxel is denoted as (N, L), and the feature P of each encoded point is shown in the formula: Step 2.3: Feature extraction. A fully connected network (FCN) layer is used to map the point features to a new feature space. A max-pooling layer is used to aggregate the point cloud features within voxels, setting the number of feature channels for each voxel to C, generating voxel feature vectors. Point cloud features are represented by a (D, H, W, C) dimensional tensor, and voxel features V are represented as follows: P′=FCN(P) V = MaxPooling(P′).

4. The point cloud 3D target detection method based on the Inception module and ASPP module according to claim 3, characterized in that: The 3D sparse backbone network in step 3 is as follows: By transforming the feature channel number to C1 using two 3D sparse residual blocks, we obtain a (D,H,W,C1)-dimensional tensor denoted as F1; then, we downsample F1 using two 3D sparse residual blocks to obtain... The tensor is denoted as F2; ​​F2 is downsampled and multi-scale features are fused using two 3D sparse Inception modules to obtain... The tensor is denoted as F3; F3 is downsampled and multi-scale features are fused using two 3D sparse Inception modules to obtain... The tensor is denoted as F4.

5. The point cloud 3D target detection method based on the Inception module and the ASPP module according to claim 4, characterized in that: Step 4 projects the 3D sparse tensor of the point cloud into two-dimensional bird's-eye view features. The specific steps are as follows: First, transpose the depth feature vector of the input 3D feature tensor. Then, perform a matrix multiplication between the transposed result and the channel feature vector to obtain a result of size [missing information]. tensor; will The tensor splicing length is The channel vector; replace the original channel vector to generate The BEV tensor.

6. The point cloud 3D target detection method based on the Inception module and the ASPP module according to claim 5, characterized in that: Step 5 uses a 2D convolutional neural network containing ASPP modules as the backbone network to extract deeper BEV features, specifically including: For the input The 3D BEV tensor is transformed using one FCN layer to increase the number of feature channels. The transformed number of feature channels is denoted as C2. The tensor is then used; next, three residual blocks are used to extract channel features, resulting in... 1D tensor; using a hollow spatial pyramid (ASPP) with 4 parallel branches to fuse multi-scale features; generating BEV tensor.

7. The point cloud 3D target detection method based on the Inception module and ASPP module according to claim 6, characterized in that: The detection network used in step 6 comprises three detection heads, each responsible for a different prediction task: a classification prediction head for identifying the target's category; a bounding box prediction head for predicting the target's bounding box; and an orientation angle prediction head for estimating the target's orientation angle. The bounding box prediction loss is denoted as L. reg The classification prediction loss is denoted as L. cls The direction angle prediction loss is denoted as L. dir The total loss is denoted as L. total ; For the detection box prediction loss L reg Parametrically convert the 3D real bounding box to (x g ,y g ,z g ,l g ,h g ,w g ,θ g The predicted anchor box is parameterized as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of the target regression 7 denoted as (x t ,y t ,z t ,l t ,h t ,w t ,θ t ), i t =θ g -θ a Where x, y, and z are the center coordinates; w, l, and h are the width, length, and height, respectively; θ is the yaw rotation about the z-axis; subscripts t, a, and g represent the encoded value, anchor value, and ground truth value, respectively; and t is the diagonal of the bottom of the anchor box; α represents the weight parameter; i represents the target index; and L is the detection box prediction loss. Reg Represented as: Classification loss L cls Expressed as a formula: L cls =-a t (1-p t ) γ log(p t ) In the formula, p t It is the estimated probability of the model, and α and γ are the parameters of the focus loss; The direction angle prediction loss is denoted as L. dir Expressed as a formula: L dir =SmoothL1(sin(θ p -θ t )) In the formula, the subscript p represents the predicted value; β1, β2, and β3 represent the weight parameters, and the total loss L total Recorded as: L total =β1L cls +β2L reg +β3L dir 。 8. The point cloud 3D target detection method based on the Inception module and ASPP module according to claim 1, characterized in that: The 3D sparse Inception module used in step 3 of the 3D sparse backbone network specifically extracts receptive field features of different sizes through four parallel 3D submanifold convolutional branches, with sizes of 1×1×1, 3×3×3, 5×5×5, and 7×7×7, respectively. The receptive fields of 5×5×5 and 7×7×7 are achieved by combining 5×1×5 with 1×5×1 and 7×1×7 with 1×7×1 submanifold sparse convolutional kernels, respectively. The features of these four branches are concatenated along the channel dimension to make the number of channels four times that of the original input. Multi-scale information fusion is performed using a 1×1×1 convolutional kernel to restore the number of channels to the original input.

Images