Three-dimensional point cloud target detection method and device based on adaptive receptive field aggregation

Abstract

The application discloses a three-dimensional point cloud target detection method and device based on adaptive receptive field aggregation, and belongs to the technical field of signal processing. The core of the method is to propose an adaptive receptive field aggregation module. In the sparse voxel feature space generated by a three-dimensional sparse convolution backbone network, the module first extracts representative features through key voxel sampling; then, based on the local density of each key voxel, the K nearest neighbor search range is dynamically adjusted to realize adaptive receptive field; further, the estimated density feature is embedded into the voxel feature and input into a multi-scale self-attention aggregation network, and according to the feature similarity, relative position and density information, adaptive aggregation weights are generated for each voxel in the neighborhood. Finally, the multi-scale feature fusion and the detection head output the target prediction box. The application significantly improves the detection accuracy of long-distance and small-scale targets, and is especially suitable for laser radar point cloud target detection in the automatic driving scene.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method and apparatus for three-dimensional point cloud target detection based on adaptive receptive field aggregation. Background Technology

[0002] 3D target detection is a core task of autonomous driving environmental perception systems. LiDAR can acquire accurate spatial depth information and is unaffected by lighting conditions. The point cloud data it generates has become a key data type for detecting targets such as vehicles and pedestrians. However, point clouds are characterized by irregularity, disorder, and highly uneven spatial distribution. In particular, the point cloud density decreases sharply with increasing detection distance, which poses a significant challenge to 3D target detection.

[0003] Currently, mainstream 3D detectors can be mainly divided into point-based, voxel-based, and point-voxel hybrid paradigms. Among them, the voxel-based method, by converting irregular point clouds into regular voxel meshes and then using 3D convolutional neural networks for feature extraction, has achieved a good balance between accuracy and efficiency and has become the most mainstream technical route. To address the high sparsity of point clouds and improve computational efficiency, 3D sparse convolutional backbone networks based on submanifold convolution are widely used, which only perform computation on non-empty voxels. However, this efficiency comes at the cost of sacrificing the receptive field. The receptive field of sparse convolution heavily depends on the distribution density of non-empty voxels in its neighborhood. In distant regions where point clouds are sparse, the effective receptive field is severely insufficient, leading to inadequate feature extraction and becoming a key bottleneck limiting detector performance, especially for distant target detection.

[0004] Existing technologies attempt to alleviate this problem through multi-scale feature fusion or adaptive sampling, but both have inherent limitations. One type of method (such as CN115082902B) employs a dynamic feature aggregation strategy based on voxel pyramids, discretely selecting and fusing features at different resolutions according to the sparsity of candidate regions. This method is essentially a cross-layer, on-off macroscopic selection, unable to achieve continuous and fine-tuning of the receptive field within a single feature layer. Another type of method (such as CN115294565A) operates in the original point cloud space, obtaining keypoints through sampling from the farthest point and learning a fixed sampling radius or optimizing the radius parameter. While this method has some adaptive capability, it operates on unstructured original point clouds, resulting in high computational costs. Furthermore, it typically learns a global, fixed aggregation radius, unable to respond to real-time density changes in local regions, and has limited ability to differentiate between sparse and dense regions.

[0005] In summary, existing technologies lack an effective means to achieve continuous, precise, and adaptive receptive field adjustment and feature aggregation based on local density information at the voxel feature space level. This is precisely the core technical problem that this invention aims to solve. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and apparatus for three-dimensional point cloud target detection based on adaptive receptive field aggregation. According to the density differences of non-empty voxels, voxel features are aggregated in an adaptive neighborhood in three-dimensional space through self-attention and cross-attention mechanisms. Low-cost multi-scale information aggregation is achieved through hierarchical search of the adaptive neighborhood.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A 3D point cloud target detection method based on adaptive receptive field aggregation, the method comprising:

[0009] Step 1: Input the original point cloud into a 3D sparse convolutional backbone network to extract sparse 3D voxel features;

[0010] Step 2: Input the sparse 3D voxel features into the bird's-eye view backbone network to generate a 3D target proposal region;

[0011] Step 3: Input the sparse 3D voxel features into the adaptive receptive field aggregation module for processing and output the enhanced key point features. This includes: sampling key voxels from the sparse 3D voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output the enhanced key point features.

[0012] Step 4: Within the proposed 3D target region, aggregate the enhanced key point features to generate proposed region features, and then output the final target prediction box through the detection head network.

[0013] Furthermore, in step 3, key voxel sampling is performed on the sparse three-dimensional voxel features. Specifically, the voxels are gradually filtered from all sparse three-dimensional voxels through two farthest point samplings. The voxel set obtained by the second sampling is the key voxel set, and the key voxel set is a proper subset of the voxel set obtained by the first sampling.

[0014] Furthermore, in step 3, the adaptive K-nearest neighbor search range is dynamically adjusted according to the local voxel density of the key voxel: when the local voxel density is lower than a preset threshold, the search range is expanded to obtain broader contextual information; when the local voxel density is higher than the preset threshold, the search range is narrowed to focus on local fine structures.

[0015] Furthermore, in step 3, the density estimation features are obtained by: calculating the average Manhattan distance from all voxels in the neighborhood voxel set to the key voxel, scaling the average value using hyperparameters, and taking the reciprocal to obtain the unnormalized density estimate.

[0016] Furthermore, in step 3, the self-attention mechanism calculates the aggregation weights based on the following three factors: the feature similarity between the key voxel and its neighboring voxels, the relative position embedding generated by the shared multilayer perceptron, and the density estimation features of the embedding.

[0017] Furthermore, in step 3, the obtained multiple aggregated features are spliced ​​together. Specifically, the self-attention mechanism with multiple different K values ​​is used to form parallel processing branches. Each branch processes and outputs aggregated features independently. The features output by all branches are spliced ​​together in the channel dimension to form multi-scale fused enhanced key point features.

[0018] Furthermore, after the self-attention mechanism aggregates the features of the neighborhood voxel set, a residual connection is introduced: the original input features of the key voxel are added to the aggregated features of the neighborhood voxel set to obtain the final output features of the key voxel.

[0019] On the other hand, the present invention provides a three-dimensional point cloud target detection device based on adaptive receptive field aggregation, comprising:

[0020] The extraction module is used to extract sparse three-dimensional voxel features from the raw point cloud input into the three-dimensional sparse convolutional backbone network.

[0021] The generation module is used to input the sparse three-dimensional voxel features into the bird's-eye view backbone network to generate a three-dimensional target suggestion region;

[0022] An adaptive module is used to input the sparse three-dimensional voxel features into an adaptive receptive field aggregation module for processing and output enhanced key point features. The module includes: sampling key voxels from the sparse three-dimensional voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output enhanced key point features.

[0023] The output module is used to aggregate the enhanced key point features within the three-dimensional target proposal area to generate proposal area features, and then output the final target prediction box through the detection head network.

[0024] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned three-dimensional point cloud target detection method based on adaptive receptive field aggregation.

[0025] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned three-dimensional point cloud target detection method based on adaptive receptive field aggregation.

[0026] The beneficial effects of this invention are as follows:

[0027] Achieving a truly local adaptive receptive field effectively alleviates the bottleneck in feature extraction in sparse regions: The adaptive receptive field aggregation module proposed in this invention can dynamically adjust the K-nearest neighbor search range based on the local point cloud density at the location of each key voxel. In sparse, distant regions of the point cloud, it automatically expands the neighborhood to obtain richer contextual information; in dense, nearby regions, it focuses on a more compact neighborhood to preserve details and reduce noise interference. This shift from a fixed receptive field to a dynamically adaptive receptive field fundamentally compensates for the insufficient receptive field of sparse convolution in distant, low-density regions.

[0028] A novel density-aware attention aggregation mechanism is proposed to improve the precision of feature fusion: This invention not only utilizes density to adjust the search range, but also embeds the estimated density features as important contextual information into voxel features and participates in the calculation of self-attention weights. This allows the feature aggregation process to simultaneously consider semantic similarity, geometric location, and local density priors, achieving soft and refined weight allocation. This mechanism can automatically suppress the contribution of irrelevant noise points in sparse regions and strengthen the features of key structures, thereby significantly improving the feature representation ability of small-scale targets (such as pedestrians and cyclists) and targets with blurred boundaries.

[0029] This invention achieves efficient multi-scale information fusion within the same feature layer, resulting in a simpler and more efficient architecture: By parallelizing multiple self-attention branches with different K values ​​within the adaptive receptive field aggregation module, this invention achieves multi-granularity information capture and fusion from local details to regional context within a single sparse voxel feature layer. This avoids the information loss and computational overhead caused by alignment and fusion across feature layers of different resolutions, providing a more direct and efficient multi-scale feature enhancement scheme. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the principle of the 3D point cloud target detection method based on adaptive receptive field aggregation of the present invention;

[0031] Figure 2 This is a schematic diagram of the principle of the multi-K nearest neighbor aggregation module of the present invention;

[0032] Figure 3a The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0033] Figure 3b for Figure 3a Point cloud corresponding to the traffic scene and target bounding box output by this invention;

[0034] Figure 3c The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0035] Figure 3d for Figure 3c Point cloud corresponding to the traffic scene and target bounding box output by this invention;

[0036] Figure 3e The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0037] Figure 3f for Figure 3e Point cloud corresponding to the traffic scene and target bounding box output by this invention;

[0038] Figure 3g The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0039] Figure 3h for Figure 3g Point cloud corresponding to the traffic scene and target bounding box output by this invention;

[0040] Figure 3i The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0041] Figure 3j for Figure 3i Point cloud corresponding to the traffic scene and target bounding box output by this invention;

[0042] Figure 3k The result of projecting an image from a real traffic scene and the target bounding box output by this invention onto the image;

[0043] Figure 3l for Figure 3kThe point cloud corresponding to the traffic scene and the target bounding box output by this invention. Detailed Implementation

[0044] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0045] like Figure 1 As shown, the present invention provides a 3D point cloud target detection method based on adaptive receptive field aggregation, which mainly includes the following steps:

[0046] Step 1: Input the original point cloud into a 3D sparse convolutional backbone network to extract sparse 3D voxel features;

[0047] First, the original irregular LiDAR point cloud is converted into a regular 3D voxel representation through spatial rasterization. Then, the voxel features containing positional encoding are input into a 3D sparse convolutional backbone network. This network, built on sparse convolution, only computes on non-empty voxels containing the point cloud, extracting 3D features layer by layer through multiple sparse convolutional modules while maintaining the sparse structure of the input data. After this process, the original point cloud is transformed into sparse 3D voxel features with rich semantic information. Each non-empty voxel contains both its discrete coordinate index in 3D space and a high-dimensional feature vector encoded by the neural network, laying the feature foundation for subsequent object detection tasks.

[0048] Step 2: Input the sparse 3D voxel features into the bird's-eye view BEV backbone network to generate 3D target proposal regions;

[0049] The sparse 3D voxel features output from step 1 are compressed and projected along the height dimension (Z-axis), generating a 2D bird's-eye view feature through methods such as max pooling or feature concatenation. This bird's-eye view feature retains the spatial distribution information and semantic features of the target on the horizontal plane, while significantly reducing the feature dimensionality. This bird's-eye view feature is then input into a BEV backbone network built on a 2D convolutional neural network. This network employs an encoder-decoder structure, generating a dense target location heatmap and a 3D bounding box parameter regression map through multi-scale feature extraction and fusion. Peak detection based on the heatmap obtains the preliminary center position of the target, and combined with the regressed size and orientation parameters, preliminary 3D target proposal regions are generated. These proposal regions provide a spatial attention mechanism for subsequent fine-grained feature extraction and bounding box optimization, effectively reducing the detection search space and improving detection efficiency.

[0050] Step 3: Input the sparse 3D voxel features into the adaptive receptive field aggregation module for processing and output the enhanced key point features. This includes: sampling key voxels from the sparse 3D voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output the enhanced key point features.

[0051] Step 3.1: Downsample the farthest point of the sparse three-dimensional voxel feature twice, wherein... This represents the sparse voxel set obtained through the first downsampling. This represents the total number of samples in the first sampling, and each element represents a voxel obtained in the first sampling. This represents the sparse voxel set obtained from the second downsampling, i.e., the key voxel set. This represents the total number of samples in the second sampling, and the element represents the voxels obtained in the second sampling, which naturally satisfy... .remember and Represent the location and feature set of a sparse 3D voxel, respectively. These represent the i-th position vector, the i-th eigenvector, and the total number of sparse 3D voxels, respectively.

[0052] Step 3.2, as follows Figure 2 As shown, for each key voxel, this method uses a K-nearest neighbor search to find a fixed number of its nearest neighbors, adaptively adjusting the search range based on the voxel density. This improves the receptive field of non-empty voxels in low-density regions. The K-neighbor voxel index numbers for a key voxel with index number m (out of a total of M) can be geometrically represented using the nearest neighbor set. It is represented as follows:

[0053] ,

[0054] The distance here is calculated using the Manhattan distance between the discrete coordinates of voxels, simplifying the original large matrix floating-point calculations into integer operations, thereby reducing the computational cost of KNN search. Based on the K-nearest neighbor search results, this method uses the voxel density estimate as the feature embedding for subsequent feature encoding. Since the K-nearest neighbor search has already been performed during feature aggregation, the density estimation based on K-nearest neighbors does not introduce any additional computational cost; only a very low additional cost is required to introduce richer contextual (density) information to the voxels. Unnormalized K-nearest neighbor-based density estimation features. It can be obtained from the following formula, where The number of nearest neighbor searches:

[0055] ,

[0056] in, It is a hyperparameter used to control scaling. Represents the nearest neighbor set, Represents the distance to the nearest neighbor. Representing probability. Obtaining density estimation features. Then, it will be embedded with the corresponding voxel features. In the middle, they are combined into an embedded feature. The set that constitutes sparse features As shown below:

[0057] ;

[0058] Step 3.3: Using a self-attention aggregation method, the weights of nearest-neighbor voxels are adaptively predicted based on the similarity, relative position, and distribution density between features. These adaptive weights are used to aggregate semantic information in the neighborhood of key voxels. In self-attention, the set of Query, Key, and Value is... It can be represented as follows:

[0059] ,

[0060] in, This represents the index as The set of Query, Key, and Value vectors corresponding to the key voxels. This indicates that the index is The set of nearest neighbors corresponding to the key voxels, where j represents an element. Relative positional relationships are used as the position embedding for Query, Key, and Value. Each key voxel group contains only K vectors, therefore only K relative offsets need to be embedded, where... To improve efficiency, a shared MLP is used to compute the location embedding for each key voxel group separately, and the location embedding feature group... It can be described as:

[0061] ,

[0062] For each element in the position P of a sparse 3D voxel, the subscript n represents the index of the position embedding.

[0063] For index number The key voxels, and the feature aggregation weights corresponding to their K nearest neighbors. The following can be calculated:

[0064] ,

[0065] The superscript T indicates transpose, and e indicates the vector embedded in the feature group.

[0066] when After the elements in the array are normalized by the Softmax function, the output features are obtained. It can be represented as:

[0067] ,

[0068] Finally, a residual path was introduced to further enhance the feature representation capability of the module. The final output of the adaptive receptive field aggregation module. It can be represented as:

[0069] ,

[0070] MLP is a multilayer perceptron network. This represents the characteristics of sparse three-dimensional voxels.

[0071] Step 3.4: Use a multi-K nearest neighbor aggregation module to aggregate features at different scales. The multi-K nearest neighbor aggregation structure implements multiple self-attention modules as described in Step 3.3, but with different K values, within the adaptive receptive field aggregation module. Finally, the features obtained from different K nearest neighbor aggregations are concatenated to form the final output features of the adaptive receptive field aggregation module. For features with a large number of nearest neighbors... of An adaptive receptive field aggregation module consists of self-attention modules, and its nearest neighbor set has... Its final output can be represented as:

[0072] ,

[0073] Among them, ARFA is the adaptive receptive field aggregation module, MKA is the multi-K nearest neighbor aggregation module, and Concat is the matrix concatenation operator. Features output by a single ARFA module This represents the final output feature of multi-scale ARFA.

[0074] Step 4: Within the proposed 3D target region, aggregate the enhanced key point features to generate proposed region features, and then output the final target prediction box through the detection head network.

[0075] Step 4.1: Aggregate the key features within the suggested region using max pooling.

[0076] Step 4.2: Perform regression and classification on the aggregated keypoint features.

[0077] To verify the effectiveness of this invention, the following simulation experiments were conducted using the KITTI dataset. The dataset includes 7481 training samples and 7518 test samples, with 3712 samples in the training set and 3769 samples in the test set. Based on the size of the target and the occlusion situation, the dataset includes three difficulty levels: Easy, Mod., and Hard. Average precision (AP) is the main evaluation metric. Point-based methods are used, Voxel-based methods are used, and PV-based methods are used. Car 3D Detection and Cyc. 3D Detection show the results of this method for the three categories of car and cyclist, respectively. The results are shown in Table 1 below.

[0078] Table 1

[0079]

[0080] Figures 3a-3l To visualize the detection results and facilitate better observation, the 3D prediction bounding box is projected onto the image. This series of images is presented in an "image-point cloud" comparison format: [Image description follows] Figure 3a , Figure 3c , Figure 3e , Figure 3g , Figure 3i , Figure 3k These represent the original camera images from different scenarios and the results of projecting the 3D detection bounding box output by this invention onto the image plane; correspondingly, Figure 3b , Figure 3d , Figure 3f , Figure 3h , Figure 3j , Figure 3l These are respectively the lidar point cloud in the same scene described above and the 3D detection box directly generated by this invention.

[0081] like Figure 3a , Figure 3c As shown, in urban road scenes containing multiple vehicles, pedestrians, and complex backgrounds, this invention can accurately detect various targets, and the detection bounding box closely matches the target contour in the two-dimensional image, with no obvious false detections or missed detections. The corresponding point cloud visualization results ( Figure 3b , Figure 3d This invention further demonstrates that it can stably output three-dimensional bounding boxes with accurate pose and reasonable size from sparse and non-uniform original point clouds, effectively restoring the true position and orientation of vehicles, pedestrians, etc. in three-dimensional space.

[0082] Figure 3e, Figure 3g , Figure 3i , Figure 3k The detection results are demonstrated under different lighting, occlusion, and distance conditions. For example, in Figure 3e / Figure 3g In the mid-to-long-range scenarios shown, despite a significant decrease in point cloud density, the present invention can still maintain reliable vehicle detection. Figure 3i / Figure 3k This demonstrates the invention's ability to distinguish between densely packed, mutually occluded targets. The corresponding point cloud results ( Figure 3f , Figure 3h , Figure 3j , Figure 3l The invention clearly demonstrates that the 3D bounding box generated by this invention can not only fit the overall shape of the point cloud cluster, but also accurately define the target in the height dimension, overcoming the inherent difficulties in height estimation of traditional bird's-eye view detection methods.

[0083] comprehensive Figures 3a-3l The series of visualization results show that the present invention can achieve stable and accurate three-dimensional target detection under different scenarios, distances, and target densities. It maintains high accuracy in two-dimensional projection and provides accurate geometric information in three-dimensional space, fully verifying the effectiveness, robustness, and practicality of the method in practical applications.

[0084] On the other hand, the present invention provides a 3D point cloud target detection device based on adaptive receptive field aggregation, which includes modules capable of implementing the steps of the aforementioned method, specifically including:

[0085] The extraction module is used to extract sparse three-dimensional voxel features from the raw point cloud input into the three-dimensional sparse convolutional backbone network.

[0086] The generation module is used to input the sparse three-dimensional voxel features into the bird's-eye view BEV backbone network to generate three-dimensional target proposal regions;

[0087] An adaptive module is used to input the sparse three-dimensional voxel features into an adaptive receptive field aggregation module for processing and output enhanced key point features. The module includes: sampling key voxels from the sparse three-dimensional voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output enhanced key point features.

[0088] The output module is used to aggregate the enhanced key point features within the three-dimensional target proposal area to generate proposal area features, and then output the final target prediction box through the detection head network.

[0089] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned three-dimensional point cloud target detection method based on adaptive receptive field aggregation.

[0090] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for three-dimensional point cloud target detection based on adaptive receptive field aggregation.

[0091] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

Claims

1. A three-dimensional point cloud target detection method based on adaptive receptive field aggregation, characterized in that, The method includes: Step 1: Input the original point cloud into a 3D sparse convolutional backbone network to extract sparse 3D voxel features; Step 2: Input the sparse 3D voxel features into the bird's-eye view backbone network to generate a 3D target proposal region; Step 3: Input the sparse 3D voxel features into the adaptive receptive field aggregation module for processing and output the enhanced key point features. This includes: sampling key voxels from the sparse 3D voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output the enhanced key point features. Step 4: Within the proposed 3D target region, aggregate the enhanced key point features to generate proposed region features, and then output the final target prediction box through the detection head network.

2. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, In step 3, key voxel sampling is performed on the sparse three-dimensional voxel features. Specifically, the voxels are gradually filtered from all sparse three-dimensional voxels through two farthest point samplings. The voxel set obtained by the second sampling is the key voxel set, and the key voxel set is a proper subset of the voxel set obtained by the first sampling.

3. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, In step 3, the adaptive K-nearest neighbor search range is dynamically adjusted according to the local voxel density of the key voxel: when the local voxel density is lower than a preset threshold, the search range is expanded to obtain broader contextual information; when the local voxel density is higher than the preset threshold, the search range is narrowed to focus on local fine structures.

4. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, In step 3, the density estimation features are obtained by: calculating the average Manhattan distance from all voxels in the neighborhood voxel set to the key voxel, scaling the average value using hyperparameters, and taking the reciprocal to obtain the unnormalized density estimate.

5. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, In step 3, the self-attention mechanism calculates the aggregation weight based on the following three factors: the feature similarity between the key voxel and its neighboring voxels, the relative position embedding generated by the shared multilayer perceptron, and the density estimation features of the embedding.

6. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, In step 3, the obtained multiple aggregated features are spliced ​​together. Specifically, the self-attention mechanism with multiple different K values ​​is used to form parallel processing branches. Each branch processes and outputs aggregated features independently. The features output by all branches are spliced ​​together in the channel dimension to form enhanced key point features with multi-scale fusion.

7. The three-dimensional point cloud target detection method based on adaptive receptive field aggregation according to claim 1, characterized in that, After the self-attention mechanism aggregates the features of the neighborhood voxel set, a residual connection is introduced: the original input features of the key voxel are added to the aggregated features of the neighborhood voxel set to obtain the final output features of the key voxel.

8. A three-dimensional point cloud target detection device based on adaptive receptive field aggregation, characterized in that, include: The extraction module is used to extract sparse 3D voxel features from the raw point cloud input into the 3D sparse convolutional backbone network. The generation module is used to input the sparse three-dimensional voxel features into the bird's-eye view backbone network to generate a three-dimensional target proposal region; An adaptive module is used to input the sparse three-dimensional voxel features into an adaptive receptive field aggregation module for processing and output enhanced key point features. The module includes: sampling key voxels from the sparse three-dimensional voxel features; for each sampled key voxel, performing an adaptive K-nearest neighbor search to obtain its neighborhood voxel set to obtain the density estimation features of the key voxel and embedding them into the corresponding key voxel features; aggregating the features of the neighborhood voxel set through a self-attention mechanism; concatenating the multiple aggregated features to output enhanced key point features. The output module is used to aggregate the enhanced key point features within the three-dimensional target proposal area to generate proposal area features, and then output the final target prediction box through the detection head network.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the three-dimensional point cloud target detection method based on adaptive receptive field aggregation as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the three-dimensional point cloud target detection method based on adaptive receptive field aggregation as described in any one of claims 1-7.

Images