Point cloud 3D target detection method based on Inception module and ASPP module

Abstract

The invention belongs to the technical field of point cloud 3D target detection, and particularly relates to a point cloud 3D target detection method based on an Inception module and an ASPP module, and the method comprises the following steps: 1, preprocessing point cloud data, cutting a point cloud, extracting a target point cloud in a region of interest, enhancing the point cloud data through employing a data enhancement technology, and obtaining the preprocessed high-quality point cloud data; the 3D sparse Inception module is integrated in the 3D sparse backbone network, and the ASPP module is introduced in the 2D dense backbone network, so that the multi-scale feature extraction capability of the network is remarkably improved; the 3D sparse Inception module is good at capturing information from multi-scale spatial features, and the ASPP module expands the receiving domain of the network and enhances the adaptability to multi-scale targets by processing convolution kernels of different sizes in parallel, which is particularly effective for improving the detection precision of small targets such as pedestrians, maintains the high detection performance of large targets, and improves the detection efficiency. And the flexibility and robustness of the model are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud 3D object detection, and in particular to a point cloud 3D object detection method based on an Inception module and an ASPP module. Background Art

[0002] As autonomous driving technology develops, the demand for accurate perception of the environment is growing. 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 the specific position and posture of the object in three-dimensional space. This is crucial to ensure the safety of autonomous vehicles, as they must understand the surrounding environment in real time to make quick and accurate decisions.

[0003] In order to improve the detection accuracy, the voxel-based 3D object detection method divides the point cloud data into three-dimensional grids (voxels), each voxel summarizes the point cloud information within its spatial area. This method uses a 3D sparse convolutional network to extract features and maps these features to a bird's-eye view (BEV) pseudo image, and finally applies 2D detection technology to identify the target.

[0004] However, current voxel-based detection methods usually use fixed-size convolution kernels to extract features. This design not only limits the model's ability to detect objects of different scales, but also hinders the effective integration of long-distance feature information. Especially when dealing with small objects 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] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention provides a point cloud 3D target detection method based on the Inception module and the ASPP module, which solves the problems raised in the above background technology.

[0007] (II) Technical solution

[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

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

[0010] Step 1: Preprocess the point cloud data, crop the point cloud, extract the target point cloud in the area of ​​interest, use data enhancement technology to enhance the point cloud data, and obtain high-quality point cloud data after preprocessing;

[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 3D sparse convolutional neural network as the backbone network, insert 3D sparse Inception module into the network, fuse multi-scale features, and output 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 the 2D convolutional neural network containing the ASPP module as the backbone network to further learn the BEV features and output a 2D dense feature tensor;

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

[0016] Furthermore, the step 1 pre-processes the three-dimensional point cloud data, crops the original point cloud data according to the (Z, Y, X) range to obtain a foreground point cloud containing the target, and enhances the point cloud data using affine transformation and GT-sampling (ground truth sampling) data enhancement technology for the cropped foreground point cloud.

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

[0018] Step 2.1: Space division, set the voxel size to (v D ,v H ,v W ), the entire point cloud space is divided into regular voxel grids, then the number of grids in the depth direction of the point cloud D = Z / v D , the number of grids in the height direction of the point cloud H = Y / v H , the number of grids in the width direction of the point cloud W = X / v W , the voxels containing point clouds are recorded as non-empty voxels, and the voxels not containing point clouds are recorded as empty voxels;

[0019] Step 2.2: Voxel encoding, set the maximum number of points of each voxel to N; for non-empty voxels with more than N points, retain N points through random sampling; for voxels with less than N points, fill them with zero vectors to N points; calculate the local mean of the point cloud in the voxel as the centroid of the voxel; encode the point cloud features in the voxel, including the coordinates of the point, reflection intensity, and the relative position of the point cloud and the centroid of the voxel, and the encoded length is recorded as L; the feature tensor of each voxel can be recorded as (N, L), and the feature P of each point after encoding is as shown in the formula:

[0020]

[0021] Step 2.3: Feature extraction, use the fully connected network (FCN) layer to map the point features to the new feature space; use the maximum pooling layer to aggregate the point cloud features within the voxel, set the number of feature channels of the voxel to C, and generate the voxel feature vector; the point cloud feature can be represented by a (D, H, W, C)-dimensional tensor, and the voxel feature V can be expressed as:

[0022] P′=FCN(P)

[0023] V = MaxPooling(P′).

[0024] Furthermore, the 3D sparse backbone network of step 3 is:

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

[0026] Furthermore, the step 4 projects the point cloud 3D sparse tensor into a two-dimensional bird's-eye view feature, and the specific steps are:

[0027] First, the depth feature vector of the input 3D feature tensor is transposed, and the result of the transposition is matrix multiplied with the channel feature vector to obtain a size of Tensor of The length of the tensor concatenation is Channel vector of ; replace the original channel vector to generate The BEV tensor.

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

[0029] For input dimensional BEV tensor, use 1 FCN layer to transform the number of feature channels, the number of feature channels after transformation is recorded as C2, and we get dimensional tensor; then use 3 residual blocks to extract channel features and get dimensional tensor; multi-scale features are fused using the Atrous Spatial Pyramid (ASPP) with 4 parallel branches; generate -dimensional BEV tensor.

[0030] Furthermore, in step 6, the detection network used includes three detection heads, which are responsible for different prediction tasks: a classification prediction head, which is used to identify the category of the target; a detection box prediction head, which is responsible for predicting the bounding box of the target; and a direction angle prediction head, which is intended to estimate the direction angle of the target. The detection box prediction loss is denoted as L reg ; The classification prediction loss is recorded as L cls ; The direction angle prediction loss is recorded as L dir ; The total loss is recorded as L total ;

[0031] For the detection box prediction loss L reg , parameterize the 3D real box as (x g ,y g ,z g ,l g ,h g ,w g ,θ g ), parameterize the predicted anchor box as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of target regression 7 Denoted as (x t ,y t ,z t ,l t ,h t ,w t ,θ t ),

[0032]

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

[0034]

[0035] Classification loss L cls The formula is:

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

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

[0038] The direction angle prediction loss is recorded as L dir The formula is:

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

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

[0041] β1, β2, β3 represent weight parameters, and the total loss L total Denoted 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 is specifically as follows: receptive field features of different sizes are extracted through four parallel 3D sub-manifold convolution branches, whose sizes are 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 respectively realized by combining sub-manifold sparse convolution kernels of 5×1×5 and 1×5×1, and 7×1×7 and 1×7×1; the features of these four branches are spliced ​​along the channel dimension to make the number of channels become 4 times that of the original input; and a 1×1×1 convolution kernel is used to perform multi-scale information fusion and restore the number of channels to the original input.

[0044] Furthermore, the main structure of the ASPP module in step 5 is as follows: The structure of the ASPP module includes a 1×1 convolution kernel and three 3×3 convolution kernels with different void rates (1, 6, and 12, respectively), and these convolutions are operated in parallel. The feature maps generated by the four parallel convolution branches are spliced ​​in the channel dimension. The number of channels of the output feature map is 4 times the number of channels of the input feature map.

[0045] (III) Beneficial effects

[0046] Compared with the prior art, the present invention provides a point cloud 3D target detection method based on the Inception module and the ASPP module, which has the following beneficial effects:

[0047] The present 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 and parameter burdens. 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 and enhances its adaptability to multi-scale targets by processing convolution kernels of different sizes in parallel. This 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the process of the present invention;

[0049] Figure 2 This is an overall network structure diagram 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 of the method of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Example

[0054] The present invention has been experimentally evaluated. The experiment divides 3712 samples from 7481 labeled samples in the KITTI dataset as a training set and 3769 samples as a test set. The overall flow chart of the present invention is as follows: Figure 1 As shown, the network structure principle diagram of the present invention is as follows Figure 2 As shown, a point cloud 3D target detection method based on an Inception module and an ASPP module proposed in one embodiment of the present invention includes the following steps:

[0055] Step 1: Data preprocessing, crop the original point cloud data along the z×y×x axis in the range of [-3,1]×[-40,40]×[0,70.4]m, and only retain the foreground point cloud data related to the target; at this time, Z=4m, Y=80m, X=70.4m; use affine transformation and GT-sampling (ground truth sampling) data enhancement technology 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: Space division, set the voxel size to (0.1m, 0.05m, 0.05m), divide the point cloud space into regular voxel grids, then the number of grids D in the depth direction of the point cloud 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; the voxels containing point clouds are recorded as non-empty voxels, and the voxels not containing point clouds are recorded as empty voxels;

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

[0059]

[0060]

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

[0062] P′=FCN(P)

[0063] V = MaxPooling(P′).

[0064] Step 3: Use the 3D sparse convolutional neural network as the backbone network, insert the 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 using two 3D sparse residual blocks to transform the number of input feature channels, the number of channels after transformation is C1 is 16, and a (40, 1600, 1408, 16)-dimensional tensor is obtained, denoted as F1; F1 is downsampled using two 3D sparse residual blocks to obtain a (20, 800, 704, 32)-dimensional tensor denoted as F2; ​​F2 is downsampled and two 3D sparse Inception modules are used to fuse multi-scale features to obtain a (10, 400, 352, 64)-dimensional tensor denoted as F3; F3 is downsampled and two 3D sparse Inception modules are used to fuse multi-scale features to obtain a (5, 200, 176, 64)-dimensional tensor denoted as F4.

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

[0067] Transpose the depth feature vector of the input 3D feature tensor F4, perform matrix multiplication on the transposed result and the channel feature vector to obtain a tensor of size (5,64); concatenate the (5,64) tensors 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: Use the 2D convolutional neural network containing the ASPP module as the backbone network to further learn the BEV features and output a 2D dense feature tensor, such as Figure 2 The 2D backbone network shown is specifically:

[0069] For the input (200, 176, 320)-dimensional BEV features, use one FCN layer to transform the number of feature channels. The number of feature channels after transformation C2 is recorded as 128, and output a (200, 176, 128)-dimensional tensor F6; use three residual blocks to process F6 to obtain a (200, 176, 128)-dimensional tensor F7; use the atrous spatial pyramid (ASPP) containing 4 parallel branches to fuse multi-scale features; generate a (200, 176, 512)-dimensional BEV tensor F8.

[0070] Step 6: Input the 2D dense feature tensor F8 into the target detection network to obtain the target detection result; specifically: use AnchorHead as the detection network, which contains three detection heads, each responsible for different prediction tasks: classification prediction head, used to identify the category of the target; detection box prediction head, responsible for predicting the bounding box of the target; direction angle prediction head, designed to estimate the direction angle of the target; classification prediction loss is recorded as L cls ; The detection box prediction loss is recorded as L reg ; The direction angle prediction head is recorded as L dir ; The total loss is recorded as L total ;

[0071] For the detection box prediction loss L reg , parameterize the 3D real box as (x g ,y g ,z g ,l g ,h g ,w g ,θ g ), parameterize the predicted anchor box as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of 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, z are the center coordinates; w, l, h are the width, length, and height respectively; θ is the yaw rotation around the z-axis; subscripts t, a, and g represent the encoded value, anchor value, and true value respectively; and is the diagonal line at the bottom of the anchor box, α represents the weight parameter, i represents the target number, and the detection box prediction loss L Reg It is expressed as:

[0075]

[0076] Classification loss L cls The formula is:

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

[0078] In the formula, p t is the estimated probability of the model, α and γ are the parameters of the focal loss, α = 0.25 and γ = 2 are used during training;

[0079] The direction angle prediction loss is recorded as Ldir The formula is:

[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 Denoted 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:

[0086] Four parallel 3D sub-manifold convolution branches are used to extract receptive field features of different sizes, which are 1×1×1, 3×3×3, 5×5×5, and 7×7×7. 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 sub-manifold sparse convolution kernels, respectively. The features of these four branches are concatenated along the channel dimension, so that the number of channels becomes four times that of the original input. Finally, a 1×1×1 convolution kernel is used to perform multi-scale information fusion and restore the number of channels to the original input.

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

[0088] The structure of the ASPP module contains a 1×1 convolution kernel and three 3×3 convolution kernels with different dilation rates (1, 6, and 12 respectively), which are operated in parallel. The feature maps generated by the four parallel convolution branches are concatenated in the channel dimension. The number of channels of the output feature map is 4 times the number of channels of the input feature map.

[0089] The point cloud detection model is trained and the model 3D detection results are output. 3D accuracy (R40) is selected as the evaluation index, and the evaluation targets include Car, Cyclist, and Pedestrian. Easy means that the minimum bounding box height of the target involved in the evaluation is 40 pixels, there is no occlusion, and the maximum truncation is 15%; Moderate means that the minimum bounding box height of the target involved in the evaluation is 25 pixels, there is partial occlusion, and the maximum truncation is 30%; Hard means that the minimum bounding box height of the target involved in the evaluation is 25 pixels, the maximum occlusion is difficult to see, and the maximum truncation is 50%. The detection results are shown in Table 1.

[0090] Table 1 KITTI dataset detection results

[0091]

[0092] From the results, it can be seen that the method described in this application shows superior performance at all difficulty levels and in all detection categories, especially in pedestrian and cyclist detection. Compared with other models such as PointPillars, Voxel-NeXt and SECOND, our model achieves high average precision of 62.56% and 64.17% respectively under medium difficulty conditions, showing effective detection capabilities for these two types of targets.

[0093] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A point cloud 3D object detection method based on Inception module and ASPP module, characterized in that: The steps include: Step 1: Preprocess the point cloud data, crop the point cloud, extract the target point cloud in the area of ​​interest, use data enhancement technology to enhance the point cloud data, and obtain high-quality point cloud data after preprocessing; 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 3D sparse convolutional neural network as the backbone network, insert 3D sparse Inception module into the network, fuse multi-scale features, and output 3D sparse tensor; Step 4: Project the 3D sparse tensor onto a 2D plane to generate Bird's Eye View (BEV) features; Step 5: Use the 2D convolutional neural network containing the ASPP module as the backbone network to further learn the BEV features and output a 2D dense feature tensor; Step 6: Input the 2D dense feature tensor into the target detection network to obtain the target detection result.

2. According to claim 1, a point cloud 3D target detection method based on Inception module and ASPP module is characterized in that: The step 1 pre-processes the three-dimensional point cloud data, crops the original point cloud data according to the (Z, Y, X) range to obtain a foreground point cloud containing the target, and enhances the point cloud data using affine transformation and GT-sampling (ground truth sampling) data enhancement technology for the cropped foreground point cloud.

3. According to claim 2, a point cloud 3D target detection method based on Inception module and ASPP module is characterized in that: The voxel feature encoding step in step 2 includes: Step 2.1: Space division, set the voxel size to (v D ,v H ,v W ), the entire point cloud space is divided into regular voxel grids, then the number of grids in the depth direction of the point cloud D = Z / v D , the number of grids in the height direction of the point cloud H = Y / v H , the number of grids in the width direction of the point cloud W = X / v W , the voxels containing point clouds are recorded as non-empty voxels, and the voxels not containing point clouds are recorded as empty voxels; Step 2.2: Voxel encoding, set the maximum number of points of each voxel to N; for non-empty voxels with more than N points, retain N points through random sampling; for voxels with less than N points, fill them with zero vectors to N points; calculate the local mean of the point cloud in the voxel as the centroid of the voxel; encode the point cloud features in the voxel, including the coordinates of the point, reflection intensity, and the relative position of the point cloud and the centroid of the voxel, and the encoded length is recorded as L; the feature tensor of each voxel can be recorded as (N, L), and the feature P of each point after encoding is as shown in the formula: Step 2.3: Feature extraction, use the fully connected network (FCN) layer to map the point features to the new feature space; use the maximum pooling layer to aggregate the point cloud features within the voxel, set the number of feature channels of the voxel to C, and generate the voxel feature vector; the point cloud feature can be represented by a (D, H, W, C)-dimensional tensor, and the voxel feature V can be expressed as: P′=FCN(P) V = MaxPooling(P′).

4. The point cloud 3D target detection method based on the Inception module and the ASPP module according to claim 3 is characterized in that: The 3D sparse backbone network of step 3 is: By using two 3D sparse residual blocks to transform the number of feature channels to C1, we get a (D, H, W, C1)-dimensional tensor denoted as F1; we downsample F1 using two 3D sparse residual blocks to get The dimensional tensor is recorded as F2; ​​F2 is downsampled and two 3D sparse Inception modules are used to fuse multi-scale features to obtain The dimensional tensor is recorded as F3; F3 is downsampled and two 3D sparse Inception modules are used to fuse multi-scale features to obtain The dimensional 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: The step 4 projects the point cloud 3D sparse tensor into a two-dimensional bird's-eye view feature, and the specific steps are: First, the depth feature vector of the input 3D feature tensor is transposed, and the result of the transposition is matrix multiplied with the channel feature vector to obtain a size of Tensor of The length of the tensor concatenation is Channel vector of ; 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: The step 5 uses a 2D convolutional neural network including an ASPP module as the backbone network to extract deeper BEV features, specifically including: For input dimensional BEV tensor, use 1 FCN layer to transform the number of feature channels, the number of feature channels after transformation is recorded as C2, and we get dimensional tensor; then use 3 residual blocks to extract channel features and get dimensional tensor; multi-scale features are fused using the Atrous Spatial Pyramid (ASPP) with 4 parallel branches; generate -dimensional BEV tensor.

7. The point cloud 3D target detection method based on the Inception module and the ASPP module according to claim 6, characterized in that: The detection network used in step 6 includes three detection heads, each of which is responsible for a different prediction task: a classification prediction head, which is used to identify the category of the target; a detection box prediction head, which is responsible for predicting the bounding box of the target; and a direction angle prediction head, which is intended to estimate the direction angle of the target. The detection box prediction loss is denoted as L reg ; The classification prediction loss is recorded as L cls ; The direction angle prediction loss is recorded as L dir ; The total loss is recorded as L total ; For the detection box prediction loss L reg , parameterize the 3D real box as (x g ,y g ,z g ,l g ,h g ,w g ,θ g ), parameterize the predicted anchor box as (x a ,y a ,z a ,l a ,h a ,w a ,θ a ), the residual vector u∈R of 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, z are the center coordinates; w, l, h are the width, length, and height respectively; θ is the yaw rotation around the z axis; subscripts t, a, and g represent the encoded value, anchor value, and true value respectively; and is the diagonal line at the bottom of the anchor box, α represents the weight parameter, i represents the target number, and the detection box prediction loss L Reg It can be expressed as: Classification loss L cls The formula is: L cls =-a t (1-p t ) γ log(p t ) In the formula, p t is the estimated probability of the model, α and γ are the parameters of the focal loss; The direction angle prediction loss is recorded as L dir The formula is: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> dir <h2 style=";text-align:left;direction:ltr"> =SmoothL1(sin(θ<h2 style=";text-align:left;direction:ltr"> p <h2 style=";text-align:left;direction:ltr"> -θ<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> )) In the formula, the subscript p represents the predicted value; β1, β2, β3 represent weight parameters, and the total loss L total Denoted as: L total =β1L cls +β2L reg +β3L dir 。 8. The point cloud 3D target detection method based on the Inception module and the ASPP module according to claim 1, characterized in that: The 3D sparse Inception module used in the 3D sparse backbone network in step 3 is specifically as follows: receptive field features of different sizes are extracted through four parallel 3D sub-manifold convolution branches, whose sizes are 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 respectively realized by combining sub-manifold sparse convolution kernels of 5×1×5 and 1×5×1, and 7×1×7 and 1×7×1; the features of these four branches are spliced ​​along the channel dimension to make the number of channels become 4 times that of the original input; and a 1×1×1 convolution kernel is used to perform multi-scale information fusion and restore the number of channels to the original input.

Images