A single-stage anchor-free 3D object detection method and system with enhanced spatial perception

Abstract

The present invention discloses a single-stage anchor-box-free 3D object detection method and system with enhanced spatial perception. A 3D feature extractor is used to extract and reduce the non-empty voxel features of point cloud data to obtain pseudo-image features. The pseudo-image features are then fused. To ensure that the extracted pseudo-image features contain more target boundary features, a foreground point segmentation branch is added. The anchor-box-free detection head regresses target parameters on the fused pseudo-image features to obtain heat maps, center point local offset, center point height position, 3D size, and orientation of each target category. The anchor-box-free detection head includes an Intersection over Union (IoU) prediction head for predicting the IoU value between the candidate bounding box and the true bounding box, and merging it into the target category confidence to obtain a final confidence score for each category. The confidence score is used to screen the heat map, center point local offset, center point height position, 3D size, and orientation. The present invention significantly improves the performance of the anchor-box-free 3D object detection algorithm.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent automobile technology, and specifically relates to a single-stage anchor-frame-free three-dimensional target detection method and system with enhanced spatial perception. Background Art

[0002] With the widespread application of LiDAR in autonomous driving and robotics, LiDAR-based 3D object detection is gaining increasing attention from both industry and academia. Compared to cameras, LiDAR offers greater robustness and adaptability to the environment, and can directly obtain depth information, making it more conducive to autonomous vehicles' perception and understanding of 3D scenes.

[0003] Based on whether or not anchor boxes are used, 3D object detection algorithms can be divided into anchor-based and anchor-free algorithms. Anchor-based algorithms can accelerate the convergence of the algorithm by pre-setting anchor boxes, but the use of anchor boxes increases the algorithm's computational cost. Anchor-free algorithms abandon the use of anchor boxes and represent the target with key points. Although this speeds up the algorithm's detection speed, it results in slow network convergence and insufficient accuracy. Based on whether or not a two-stage correction module is used, 3D object detection algorithms can be further divided into single-stage and two-stage categories. Single-stage detection algorithms are fast but have low accuracy, while two-stage algorithms have high accuracy but are slow. Considering the real-time requirements of algorithms in autonomous vehicles, single-stage anchor-free 3D object detection algorithms with simpler structures are increasingly attracting attention from academia and industry.

[0004] The sparsity of point cloud data has always been a challenge for 3D object detection. Due to the limitations of lidar and factors such as occlusion, the number of point clouds assigned to each target is extremely limited. This requires the 3D detection network to possess strong spatial perception capabilities, capable of predicting the global size of the target based on local point cloud data. Without the spatial prior information provided by anchor boxes, current anchor-free detection networks lack sufficient spatial perception capabilities, making them prone to missed detections and false detections of targets with small point cloud counts. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a single-stage anchor-free 3D target detection method and system with enhanced spatial perception, which improves the spatial perception capability of the 3D detection network and thus enhances the performance of the anchor-free 3D target detection algorithm.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] A single-stage anchor-free 3D object detection method with enhanced spatial perception:

[0008] The three-dimensional feature extractor is used to extract the non-empty voxel features of the point cloud data and perform dimensionality reduction processing to obtain pseudo image features;

[0009] The pseudo image feature fusion module fuses the pseudo image features;

[0010] The anchor-free detection head regresses the target parameters of the fused pseudo-image features to obtain the heat map, center point local offset, center point height position, 3D size and orientation of each category of targets;

[0011] The three-dimensional feature extractor consists of four stages. The first stage consists of submanifold convolution and two spatial residual modules. The second, third and fourth stages each consist of sparse convolution and two spatial residual modules. The number of feature channels in each stage is 32, 64, 128 and 128 respectively.

[0012] The spatial group of the pseudo image feature fusion module inputs pseudo image features and outputs spatial geometric features. The spatial geometric features are subjected to feature channel transformation, and the semantic group performs downsampling operation and feature channel transformation on the spatial geometric features; the semantic features after upsampling operation and feature channel transformation are superimposed with the spatial geometric features after feature channel transformation to obtain semantically enhanced spatial geometric features; channel transformation and size transformation are performed on the semantically enhanced spatial geometric features and the semantic features after upsampling operation and feature channel transformation respectively, and the semantically enhanced spatial geometric features and semantic features after channel transformation and size transformation are weighted and superimposed.

[0013] Furthermore, the anchor-free box detection head includes an IoU prediction head, which is used to predict the IoU value between the candidate bounding box and the true bounding box, and merge it into the target category confidence to obtain the final confidence score of each category. The confidence score is used to screen the heat map, center point local offset, center point height position, three-dimensional size and orientation to obtain the final heat map, center point local offset, center point height position, three-dimensional size and orientation of the three-dimensional target.

[0014] Furthermore, the target boundary features in the pseudo image features are enhanced using the foreground point segmentation branch:

[0015] The foreground point segmentation is performed using the fusion point feature, and the formula of the fusion point feature is:

[0016]

[0017]

[0018] in: is the fusion point feature, w j (p i ) is the distance between the K feature points of the original point cloud and the original point in the ONCE dataset, Represents the nearest neighbor area, ε is a minimum positive value, pi is the coordinate of the point in the ONCE dataset, p j are the feature point coordinates corresponding to the non-empty voxel features at each stage of the 3D feature extractor, and M represents the total number of feature points.

[0019] Furthermore, the spatial group and the semantic group are both composed of one convolutional layer and two self-correction convolutional layers, the number of feature channels of the spatial group is 128, and the number of feature channels of the semantic group is 256.

[0020] Furthermore, the spatial residual module consists of three parts. The first part is connected in sequence by submanifold convolution, batch normalization layer, and ReLU activation layer, and the third part is connected in sequence by submanifold convolution, batch normalization layer, and SE-layer modules.

[0021] Furthermore, the non-empty voxel features are obtained by voxelizing the pre-processed point cloud data.

[0022] Furthermore, the weighted processing is implemented using an attention map, which includes a layer of convolution to predict spatial features and semantic features.

[0023] A single-stage anchor-free 3D object detection system with enhanced spatial perception, including:

[0024] Voxel processing module, which converts the point cloud data in the ONCE dataset into voxel-level features;

[0025] 3D feature extractor, which introduces the spatial attention mechanism SE-layer module and residual structure to extract pseudo-image features;

[0026] Pseudo-image feature fusion module, which uses a spatial semantic feature aggregation architecture to fuse semantic features and spatial geometric features;

[0027] Anchor-free detection head for final target category regression and bounding box prediction;

[0028] The foreground point segmentation branch assists the 3D feature extractor to learn more target boundary features.

[0029] An electronic device comprising a memory and a processor;

[0030] The memory is used to store computer programs;

[0031] The processor is used to execute the computer program and implement the above-mentioned single-stage anchor-free three-dimensional object detection method when executing the computer program.

[0032] A storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the above-mentioned single-stage anchor-free three-dimensional object detection method.

[0033] The beneficial effects of the present invention are:

[0034] (1) By constructing a more effective three-dimensional feature extractor and a pseudo-image feature fusion module, the present invention can extract more refined spatial geometric features, and the fused features can better meet the detection of various categories of targets;

[0035] (2) By adding a lightweight IoU prediction head to the anchor-free detection head, the accuracy of bounding box size prediction is further improved, and the mismatch between target category confidence and location confidence is effectively alleviated;

[0036] (3) By adding a foreground point segmentation branch, the 3D feature extractor is prompted to extract more target boundary features, further enhancing the algorithm's spatial perception ability and significantly reducing false detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flowchart of the single-stage anchor-free 3D object detection with enhanced spatial perception according to the present invention;

[0038] Figure 2 This is a network structure diagram of the three-dimensional feature extractor of the present invention;

[0039] Figure 3 This is a structural diagram of the pseudo image feature fusion module of the present invention;

[0040] Figure 4 This is the self-correcting convolution structure diagram of the present invention;

[0041] Figure 5 This is a diagram of the foreground point segmentation branch structure of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0043] like Figure 1 As shown, the present invention provides a single-stage anchor-free 3D target detection method with enhanced spatial perception, which specifically includes the following steps:

[0044] Step 1: Point cloud data preprocessing

[0045] The true value of the annotated point cloud data in the ONCE dataset is extracted to establish a true value sample database. The present invention uses different data enhancement methods to preprocess the point cloud data in the ONCE dataset. The data enhancement methods used in the present invention include: (1) true value sampling, which selects a number of true value samples for each category target from the true value sample database and directly puts them into the point cloud data of the training set. The categories include cars, pedestrians, trucks, riders and buses; (2) random flipping operation, which randomly flips the point cloud data obtained in (1) along the X-axis and Y-axis of the laser radar coordinate system, and the flip angle range is set to (3) Random global scaling, the point cloud data obtained in (2) is Performs random global scaling within the range.

[0046] Step 2: Voxelization of point cloud

[0047] Point cloud voxelization is performed by the voxelization processing module. The voxelization processing module inputs pre-processed point cloud data of a shape of B×N×4, where B and N represent the number of point cloud frames and the number of points in each frame of the scene, respectively. Each point in the input point cloud data consists of its three-dimensional coordinates (x, y, z) and reflection intensity r in the lidar coordinate system. The point cloud is divided into voxels along the X, Y, and Z axes of the lidar coordinate system according to the predefined voxel size; then, each point is assigned to the corresponding voxel according to its three-dimensional coordinates; finally, the average feature value of all points belonging to the same voxel is calculated to obtain the three-dimensional voxel-level features of the point cloud data, which facilitates the efficient processing of voxel features by three-dimensional convolution. The specific structure of the voxelization processing module is existing technology and will not be repeated here.

[0048] Step 3: 3D feature extraction

[0049] The three-dimensional feature extraction is completed by processing the non-empty voxel features through the three-dimensional feature extractor, and the voxel-level features of the point cloud data are queried using a hash table to obtain the non-empty voxel features. In order to meet the needs of multi-target detection in autonomous driving scenarios, the features extracted by the three-dimensional feature extractor need to fully meet the detection of targets of various categories; however, the existing three-dimensional feature extractors used for anchor-free detection algorithms are not only difficult to extract sufficiently fine spatial features, but also the extracted spatial features are difficult to balance the detection of targets of various categories. In this regard, this embodiment constructs a more effective three-dimensional feature extractor, such as Figure 2As shown in the figure, the 3D feature extractor consists of four stages. The first stage consists of submanifold convolution and two spatial residual modules. The second, third, and fourth stages each consist of sparse convolution and two spatial residual modules. The number of feature channels in each stage is 32, 64, 128, and 128, respectively. For the first stage, submanifold convolution is first used for feature extraction, followed by downsampling with a stride of 1, followed by feature extraction using two spatial residual modules. The spatial residual module consists of three parts: the first part is composed of a submanifold convolution, a batch normalization layer, and a ReLU activation layer connected in sequence, the second part is composed of a submanifold convolution, a batch normalization layer, and an SE-layer module connected in sequence, and the third part is a ReLU activation layer. The SE-layer module introduces a spatial attention mechanism that focuses on the connection between local and global features, thereby extracting more refined spatial features that are more suitable for detecting various object categories. Furthermore, to reduce information loss during feature extraction, a residual structure is introduced, superimposing the input of the spatial residual module with the output of the SE-layer module. Finally, the superimposed features are processed in a ReLU activation layer. Sparse convolution is used for feature extraction and downsampling, with the second, third, and fourth stages performing downsampling with strides of 2, 4, and 8, respectively. After four stages of feature extraction and dimensionality reduction along the height direction, a pseudo-image feature of size 256 × H × W is obtained, where H and W are the height and width of the pseudo-image, respectively.

[0050] Step 4: Pseudo image feature fusion

[0051] Pseudo image feature fusion is performed by the pseudo image fusion module. The pseudo image feature fusion module further fuses the pseudo image features, and the fused features will be used for the final target classification and bounding box regression. High-level semantic features are conducive to target classification, while shallow spatial geometric features are very important for the regression of the target bounding box. Therefore, when the pseudo image feature fusion module performs feature fusion, it is very critical to balance the spatial geometric features and semantic features. However, the pseudo image feature extractor of the existing anchor-free frame detection algorithm cannot efficiently balance the spatial geometric features and semantic features. In order to solve this problem, this embodiment designs a more effective pseudo image feature fusion module, such as Figure 3As shown in the figure, this module uses a spatial semantic aggregation architecture that can adaptively fuse high-level, abstract semantic features with low-level spatial geometric features. Specifically, the spatial group of the pseudo-image feature fusion module inputs pseudo-image features and outputs coarse-grained spatial geometric features, which are then passed to the semantic group. Both the spatial group and the semantic group consist of one convolutional layer and two self-corrected convolutional layers. The spatial group has 128 feature channels, while the semantic group has 256 feature channels. To avoid excessive loss of spatial geometric features and ensure consistency between input and output sizes, the semantic group then takes the spatial geometric features extracted by the spatial group as input and performs a downsampling operation with a stride of 2, simultaneously reducing the number of feature channels from 128 to 256, thereby outputting refined, high-level semantic features. To obtain spatial geometric features that better distinguish between different categories of objects, the semantic features output by the semantic group are upsampled and subjected to feature channel transformation. The spatial geometric features output by the spatial group are also subjected to feature channel transformation. The semantic features after upsampling and feature channel transformation are superimposed with the spatial geometric features after feature channel transformation to obtain semantically enhanced spatial geometric features. The semantically enhanced spatial geometric features, upsampling operations, and semantic features after feature channel transformation are subjected to channel transformation and size transformation respectively to ensure the consistency of feature channels and sizes of the two. Finally, in order to adaptively balance the spatial geometric features and semantic features, a layer of convolution is used to predict the attention maps of spatial features and semantic features respectively, and the attention maps are used to weight the semantically enhanced spatial geometric features and semantic features after channel transformation and size transformation respectively, and then the weighted semantically enhanced spatial geometric features and semantic features are superimposed. In order to make the fused features more able to meet the detection of various categories of targets, the pseudo image feature fusion module uses self-correction convolution to replace ordinary convolution, such as Figure 4 As shown, the self-correcting convolution introduces channel and spatial attention mechanisms, which not only expands the network's receptive field but also enables the pseudo-image fusion module to better capture the connection between local and global features. The pseudo-image fusion module of this embodiment better balances spatial geometric and semantic features, and the fused features are more suitable for detecting various targets.

[0052] Step 5: Regression of target parameters

[0053] The regression of target-related parameters is performed by the anchor-free detection head. In order to more efficiently and accurately regress the target bounding box parameters and predict the target category confidence, this embodiment uses the anchor-free detection head to regress the target parameters on the fused pseudo-image features. The anchor-free box detection head consists of six sub-prediction heads. In addition to the commonly used heat map prediction head, center point local offset prediction head, center point height position prediction head, three-dimensional size prediction head and orientation prediction head, this embodiment adds a lightweight IoU prediction head; the heat map prediction head, center point local offset prediction head, center point height position prediction head, three-dimensional size prediction head and orientation prediction head predict the fused pseudo image features, and obtain the heat map, center point local offset, center point height position, three-dimensional size and orientation of each category of targets respectively; the IoU prediction head is used to predict the IoU value between the candidate bounding box and the true bounding box, which can be used as a metric for target positioning confidence. In this embodiment, the IoU value predicted by the IoU prediction head is merged into the target category confidence, and a post-processing function is used to recalculate the final confidence score CF of each category:

[0054] CF=score 1-α *iou α (1)

[0055] Where: score is the initial category confidence score, iou is the predicted IoU value, α is a hyperparameter ∈ [0, 1], which is used to control the ratio between the category confidence score and the predicted IoU value; the confidence score CF is used to screen the heat map, center point local offset, center point height position, three-dimensional size and orientation to obtain the final heat map, center point local offset, center point height position, three-dimensional size and orientation of the three-dimensional target.

[0056] The added IoU prediction head effectively enhances the detection system's perception of the size of targets of each category, and alleviates the mismatch between target category confidence and location confidence.

[0057] Step 6: Auxiliary training of foreground point segmentation

[0058] The foreground point segmentation auxiliary training is completed by the foreground point segmentation branch. In point cloud-based 3D target detection, accurately distinguishing the foreground and background can effectively reduce missed detections and false detections, while the perception of the target boundary can promote the distinction of the foreground from the background. In order to enable the 3D feature extractor to learn more refined target boundary features (target boundary features are included in the pseudo-image features), the non-empty voxel feature coordinates corresponding to the four stages of the 3D feature extractor are converted into real coordinates. According to the downsampling step size of the current stage, the non-empty voxel features of each stage correspond to a feature point coordinate, and {(f j ,p j):j=1,…,M} represents the feature point, f j is the eigenvector, p j is the feature point coordinate, M represents the total number of feature points; the calculation formula of the feature point coordinate is as follows:

[0059] p j =(VC j +0.5)*VS*DT+R (2)

[0060] Among them: VC represents the coordinates of the center point of the non-empty voxel, VS represents the size of the non-empty voxel, DT is the downsampling multiple, and R represents the point cloud range.

[0061] Then, the features extracted by the four stages of the 3D feature extractor are interpolated to the positions of the point cloud in the ONCE dataset using the interpolation method. The point coordinates in the ONCE dataset are represented as {p i :i=1,…,N},p i is the point coordinate, and N is the total number of points. For interpolation, it is necessary to find the K (in this embodiment, K is 3) nearest neighbor feature points of each point in the original point cloud and calculate the distance w between the K feature points and the original point. j (p i ), then the inverse distance weighted average is used for the K nearest neighbor feature points to obtain the interpolation features at the original point cloud position, and finally the interpolation features from different stages at the same position in the original point cloud are spliced ​​together to obtain the final fusion point feature for foreground point segmentation, which is expressed as Each fusion point feature The calculation formula is as follows:

[0062]

[0063]

[0064] in: Represents the nearest neighbor area; ε is a minimum positive value, and in this embodiment, it is 10 -8 .

[0065] Points within the true bounding box are considered foreground points. In order to distinguish whether they are foreground or background, the foreground points are segmented by fusing point features. The foreground point segmentation branch effectively helps the 3D feature extractor extract pseudo-image features containing more target boundary features, further improving spatial feature perception capabilities without affecting the algorithm's calculation speed.

[0066] Step 7: Model training and loss function setting

[0067] Different sub-prediction heads in the anchor-free box detection head use different loss functions. This embodiment uses Focalloss to supervise the training of the heat map prediction head, while L1 loss is used for supervised training of the center point local offset prediction head, center point height position prediction head, 3D size prediction head, orientation prediction head, and IoU prediction head. Focal loss is also used for supervised training of the foreground point segmentation branch.

[0068] The weighted sum of all losses is taken as the final loss:

[0069] L box =λ off L off +λ z L Z +λ size L size +λ ori L ori (5)

[0070] L=λ heat L heat +L box +λ iou L iou +λ seg L seg (6)

[0071] Among them L heat , L off , L Z , L size , L ori , L iou , L seg The losses corresponding to the heat map prediction head, the center local offset prediction head, the center height position prediction head, the 3D size prediction head, the orientation prediction head, the IoU prediction head and the foreground point segmentation branch are λ respectively. heat ,λ off ,λ z ,λ size ,λ ori ,λ iou ,λ seg are the weights of the corresponding losses of the heat map prediction head, the local offset prediction head of the center, the height position prediction head of the center, the three-dimensional size prediction head, the orientation prediction head, the IoU prediction head and the foreground point segmentation branch, respectively. box An intermediate amount.

[0072] A single-stage anchor-free 3D object detection system with enhanced spatial perception, including:

[0073] The voxel processing module converts the irregular point cloud data in the ONCE dataset into regular 3D voxel-level features. The voxelized point cloud data can be efficiently processed by 3D convolution.

[0074] 3D feature extractor, which introduces the spatial attention mechanism SE-layer module and residual structure to extract pseudo-image features;

[0075] The pseudo-image feature fusion module uses a spatial semantic feature aggregation architecture to adaptively fuse high-level abstract semantic features with low-level spatial geometric features;

[0076] Anchor-free detection head for final target category regression and bounding box prediction;

[0077] The foreground point segmentation branch assists the 3D feature extractor to learn more target boundary features.

[0078] Based on the same inventive concept as a single-stage anchor-box-free 3D object detection method with spatial awareness enhancement, the present application also provides an electronic device comprising one or more processors and one or more memories, wherein the memories store computer-readable code. When executed by the one or more processors, the computer-readable code implements the single-stage anchor-box-free 3D object detection method with spatial awareness enhancement. The memories may include a non-volatile storage medium and internal memory; the non-volatile storage medium may store an operating system and the computer-readable code. The computer-readable code includes program instructions that, when executed, cause the processor to perform any single-stage anchor-box-free 3D object detection method with spatial awareness enhancement. The processor is configured to provide computing and control capabilities, supporting the operation of the entire electronic device. The memory provides an environment for the execution of the computer-readable code in the non-volatile storage medium. When executed by the processor, the computer-readable code causes the processor to perform any single-stage anchor-box-free 3D object detection method with spatial awareness enhancement.

[0079] It should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0080] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer-readable code, and the computer-readable code includes program instructions. The processor executes the program instructions to implement the single-stage anchor-free three-dimensional target detection method with enhanced spatial perception of the present application.

[0081] The computer-readable storage medium may be an internal storage unit of the electronic device described in the aforementioned embodiment, such as a hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital (SD) card, a flash memory card, etc. equipped on the electronic device.

[0082] The embodiments described are preferred implementations of the present invention, but the present invention is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. A single-stage anchor-free 3D object detection method with enhanced spatial perception, characterized by: The three-dimensional feature extractor is used to extract the non-empty voxel features of the point cloud data and perform dimensionality reduction processing to obtain pseudo image features; The pseudo image feature fusion module fuses the pseudo image features; The anchor-free detection head regresses the target parameters of the fused pseudo-image features to obtain the heat map, center point local offset, center point height position, 3D size and orientation of each category of targets; The three-dimensional feature extractor consists of four stages. The first stage consists of submanifold convolution and two spatial residual modules. The second, third and fourth stages each consist of sparse convolution and two spatial residual modules. The number of feature channels in each stage is 32, 64, 128 and 128 respectively. The spatial group of the pseudo image feature fusion module inputs pseudo image features and outputs spatial geometric features, the spatial geometric features are subjected to feature channel transformation, and the semantic group performs downsampling and feature channel transformation on the spatial geometric features; the semantic features after upsampling and feature channel transformation are superimposed with the spatial geometric features after feature channel transformation to obtain semantically enhanced spatial geometric features; Perform channel transformation and size transformation on the semantically enhanced spatial geometric features, upsampling operations, and semantic features after feature channel transformation, respectively, and perform weighted processing on the semantically enhanced spatial geometric features and semantic features after channel transformation and size transformation and superimpose them; The anchor-free box detection head includes an IoU prediction head, which is used to predict the IoU value between the candidate bounding box and the true bounding box, and merge it into the target category confidence to obtain the final confidence score for each category. The confidence score is used to screen the heat map, center point local offset, center point height position, three-dimensional size and orientation to obtain the final heat map, center point local offset, center point height position, three-dimensional size and orientation of the three-dimensional target; The target boundary features in the pseudo image features are enhanced by using the foreground point segmentation branch; The foreground point segmentation is performed using the fusion point feature, and the formula of the fusion point feature is: , ,in: is the fusion point feature, is the distance between the K feature points of the original point cloud and the original point in the ONCE dataset, represents the nearest neighbor region, is a very small positive value. is the coordinate of the point in the ONCE dataset, is the feature point coordinate corresponding to the non-empty voxel feature of each stage of the 3D feature extractor, M represents the total number of feature points, is the feature vector.

2. The single-stage anchor-free 3D object detection method according to claim 1, characterized in that: The spatial group and the semantic group are both composed of one convolutional layer and two self-correction convolutional layers. The number of feature channels of the spatial group is 128, and the number of feature channels of the semantic group is 256.

3. The single-stage anchor-free 3D object detection method according to claim 1, characterized in that: The spatial residual module consists of three parts. The first part is connected in sequence by submanifold convolution, batch normalization layer, and ReLU activation layer, and the third part is the ReLU activation layer.

4. The single-stage anchor-free 3D object detection method according to claim 1, characterized in that: The non-empty voxel features are obtained by voxelizing the pre-processed point cloud data.

5. The single-stage anchor-free 3D object detection method according to claim 1, characterized in that: The weighted processing is implemented using an attention map, which contains a layer of convolution to predict spatial features and semantic features.

6. A system for implementing the single-stage anchor-free 3D object detection method according to any one of claims 1 to 5, characterized in that: include: Voxel processing module, which converts the point cloud data in the ONCE dataset into voxel-level features; 3D feature extractor, which introduces the spatial attention mechanism SE-layer module and residual structure to extract pseudo-image features; Pseudo-image feature fusion module, which uses a spatial semantic feature aggregation architecture to fuse semantic features and spatial geometric features; Anchor-free detection head for final target category regression and bounding box prediction; The foreground point segmentation branch assists the 3D feature extractor to learn more target boundary features.

7. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program and implement the single-stage anchor-free three-dimensional object detection method according to any one of claims 1 to 6 when executing the computer program.

8. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the single-stage anchor-free three-dimensional object detection method according to any one of claims 1 to 6.

Images