A target detection method and system based on a 4D laser radar, a computer device, and a storage medium

Abstract

A target detection method based on 4D LiDAR includes: acquiring a sparse point cloud output from a 4D LiDAR, inputting it into a hybrid voxel cylinder encoder for processing, wherein the hybrid voxel cylinder encoder processing method for the sparse point cloud includes: voxelizing the sparse point cloud to obtain an original feature map, generating voxel features and voxel coordinates; performing channel compression on the voxel features to obtain voxel descriptors, generating a concise feature map based on all voxel descriptors; reconstructing the concise feature map based on the voxel coordinates to obtain an optimized feature map; processing the optimized feature map using a pre-trained state space model in several stages to obtain several basic feature maps; fusing all basic feature maps to obtain a multi-scale fused feature map; and detecting the multi-scale fused feature map to obtain a target detection result. This invention provides a lightweight, high-precision target detection method, system, computer device, and storage medium based on 4D LiDAR.

Description

Technical Field

[0001] This invention relates to the field of lidar target detection technology, specifically a target detection method, system, computer equipment, and storage medium based on 4D lidar. Background Technology

[0002] One of the core challenges of autonomous driving is reliably perceiving the environment in complex and dynamic scenarios, including dealing with inclement weather and varying lighting conditions. Existing visual sensors experience a sharp decline in performance under low light or inclement weather conditions; while LiDAR, although providing high-resolution 3D point clouds, is limited by its low-wavelength signals in adverse weather conditions such as rain and snow. In contrast, 4D LiDAR maintains high robustness even in inclement weather, making it a highly attractive option for all-weather perception.

[0003] Existing public datasets, such as K-Radar, Astyx, VoD, and TJ 4DRadSet, typically use radar sensors operating at frame rates of 10-15 Hz. Therefore, real-time object detection systems generally require a minimum inference speed of 15-20 FPS. Researchers have found that 4D LiDAR extends traditional 3D radar by increasing the elevation angle, achieving a denser 3D spatial representation. This richer information enables reliable object shape inference and more accurate localization, making it suitable for higher levels of autonomous driving. Existing work has achieved an extremely high inference speed of 34.3 FPS on high-end edge devices using techniques such as velocity decomposition and pillar attention. Furthermore, existing models often rely on computationally intensive voxels or Transformer-based backbone networks, whose computational complexity leads to severe consumption of memory and computing resources, making real-time detection difficult on resource-constrained edge platforms. Therefore, deploying deep learning models on edge computing platforms to meet the stringent real-time requirements of autonomous driving remains a significant challenge.

[0004] Furthermore, state-space models, which have garnered significant attention in the past two years, have achieved remarkable success in the field of 3D point cloud perception. The recently proposed original Mamba model differs from the traditional Transformer model in several ways. With varying levels of complexity, Mamba implements... The linear inference time of LiDAR makes it extremely suitable for processing long sequences of features on resource-constrained hardware. Current state-of-the-art 3D target detection models based on LiDAR, such as VoxelMamba and PillarMamba, utilize the original Mamba architecture to model the spatial / temporal dependencies in voxel and pillar features. While this improves detection accuracy, they are primarily designed for high-performance computing environments, resulting in large model sizes and high computational budgets, which limits their applicability to resource-constrained embedded platforms. Directly applying LiDAR-designed architectures to 4D LiDAR data leads to a significant decrease in model accuracy due to the sparse, noisy, and irregular nature of LiDAR data.

[0005] In summary, overcoming the shortcomings of 4D LiDAR data sparseness and high noise, while breaking through the bottleneck of high computational complexity of traditional network architecture, and designing a target detection system that can guarantee high accuracy in harsh weather conditions and low latency real-time operation on low-computing-power edge devices, is a key technical challenge that urgently needs to be addressed in this field. Summary of the Invention

[0006] This invention aims to solve the technical problems of high computational complexity, difficulty in real-time deployment on resource-constrained edge devices, and insufficient accuracy of traditional models when processing sparse radar point clouds in existing 4D radar target detection systems. It provides a lightweight, high-precision target detection method, system, computer equipment, and storage medium based on 4D radar.

[0007] To achieve the above objectives, the specific solution adopted by the present invention is as follows: a target detection method based on 4D lidar, comprising: The sparse point cloud output from the 4D LiDAR is acquired and input into a hybrid voxel cylinder encoder for processing. The processing method of the hybrid voxel cylinder encoder for the sparse point cloud includes: The sparse point cloud is voxelized to obtain the original feature map, and voxel features and voxel coordinates are generated. Voxel features are channel compressed to obtain voxel descriptors, and a concise feature map is generated based on all voxel descriptors; An optimized feature map is obtained by reconstructing a simplified feature map based on voxel coordinates; The process involves using a pre-trained state-space model to process the optimized feature map in several stages to obtain several basic feature maps, fusing all the basic feature maps to obtain a multi-scale fused feature map, and then performing detection on the multi-scale fused feature map to obtain the target detection result. The state-space model processing methods in each stage include: The input data is flattened to obtain a one-dimensional token sequence and fed into the first and second processing paths in parallel. The input data is either an optimized feature map or a basic feature map output from the previous stage. In the first processing path, the channel dimension of the one-dimensional token sequence is increased by feature projection, and in the second processing path, the channel dimension of the one-dimensional token sequence is reduced by feature compression. The basic feature map is obtained by back-projecting and feature fusion of the first processing result of the first processing path and the second processing result of the second processing path.

[0008] As a further optimization of the above-mentioned target detection method based on 4D LiDAR, the method for obtaining the original feature map by voxelizing sparse point clouds includes: Regions of interest (ROIs) are obtained by cropping sparse point clouds. The region of interest is divided into multiple regular 3D voxel grids to obtain the original feature map.

[0009] As a further optimization of the above-mentioned target detection method based on 4D LiDAR: when the region of interest is divided into multiple regular 3D voxel grids, the voxel features of each non-empty regular 3D voxel grid are generated, and the discrete spatial index information of each non-empty regular 3D voxel grid is recorded to obtain the voxel coordinates.

[0010] As a further optimization of the above-mentioned target detection method based on 4D LiDAR, the method for obtaining voxel descriptors by channel compression of voxel features includes sequential linear projection processing, max pooling processing, and layer normalization processing.

[0011] As a further optimization of the above-mentioned target detection method based on 4D LiDAR: when reconstructing the simplified feature map based on voxel coordinates, the PillarScatter3D module is used to reconstruct the optimized feature map.

[0012] As a further optimization of the above-mentioned target detection method based on 4D LiDAR: during the process of the state space model processing the optimized feature map, the input data is downsampled and convolved before any stage.

[0013] As a further optimization of the above-mentioned target detection method based on 4D LiDAR: when fusing all basic feature maps, channel stitching is performed on all basic feature maps to form a multi-scale fused feature map.

[0014] A target detection system based on 4D LiDAR, used to implement the above-mentioned target detection method based on 4D LiDAR, the system comprising: The data acquisition module is used to acquire sparse point clouds from 4D LiDAR; The point cloud processing module is used to process sparse point clouds using a hybrid voxel cylinder encoder to obtain optimized feature maps. The object detection module is used to process the optimized feature map in several stages using a pre-trained state space model to obtain several basic feature maps, fuse all the basic feature maps to obtain a multi-scale fused feature map, and detect the multi-scale fused feature map to obtain the object detection result.

[0015] Computer equipment, including: Memory, used to store computer programs; A processor is used to read and execute the computer program to implement the above-described target detection method based on 4D LiDAR.

[0016] A storage medium for storing a computer program that, when executed, implements the aforementioned target detection method based on 4D LiDAR.

[0017] Beneficial effects: This invention proposes a hybrid voxel cylinder encoder for 4D LiDAR, which effectively preserves vertical structure information and improves the richness of feature representation during the mapping of sparse point clouds to two-dimensional features. The backbone network structure based on the state space model designed in this invention abandons the traditional high-energy-consuming attention mechanism computation, and achieves a perfect combination of long-distance feature capture and linear time complexity, which significantly reduces the number of model parameters and memory usage. This invention balances high accuracy and lightweight design, and can achieve efficient real-time inference speed on edge computing devices. At the same time, it can exhibit robustness and target detection accuracy far exceeding traditional baseline models under various adverse weather conditions such as rain, snow, and fog. Attached Figure Description

[0018] Figure 1 This is an overall architecture diagram of a high-efficiency 4D lidar target detection system for edge devices provided by the present invention; Figure 2 This is a flowchart illustrating the internal structure and processing flow of the hybrid voxel column encoder in this embodiment; Figure 3 This is a detailed structural diagram of the backbone network in this embodiment; Figure 4 This is a comparison diagram of the inference time decomposition of the present invention and the benchmark model in this embodiment. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1 to 3As shown, a target detection method based on 4D lidar includes S1 to S2.

[0021] S1. Obtain the sparse point cloud output from the 4D LiDAR and input it into a hybrid voxel cylinder encoder for processing. In this invention, each point in the sparse point cloud output by the 4D LiDAR contains four core attributes and is labeled (X, Y, Z, PW), where (X, Y, Z) are the three-dimensional spatial coordinates of the point, and PW is the radar power-related feature. Considering the characteristics of sparse point clouds, such as sparsity, high noise, and irregular distribution, this invention employs a hybrid voxel cylinder encoder. During the process of mapping the sparse point cloud to low-dimensional features, the vertical structure information is effectively preserved and the computational complexity of subsequent processing is reduced through a collaborative mechanism of voxelization feature extraction and cylinderization spatial reconstruction.

[0022] Specifically, the hybrid voxel cylinder encoder's processing method for sparse point clouds includes S11 to S13.

[0023] S11. The sparse point cloud is voxelized to obtain the original feature map, and voxel features and voxel coordinates are generated. Specifically, the method for voxelizing the sparse point cloud to obtain the original feature map includes S111 to S112.

[0024] S111. The sparse point cloud is cropped to obtain the Region of Interest (ROI). The specific method for ROI cropping is existing technology in this field and will not be elaborated here. The range of ROI cropping is set according to the actual application scenario. For example, in one embodiment of this invention, the invention is applied to an autonomous driving scenario. In this case, a certain range in front of the vehicle is designated as the ROI. The ROI can be a spatial area of ​​100 meters forward, 50 meters laterally, and 5 meters high, to eliminate interference from distant irrelevant point clouds and reduce the computational burden of subsequent voxelization processing.

[0025] S112. The region of interest is divided into multiple regular 3D voxel grids to obtain the original feature map. More specifically, when dividing the region of interest into multiple regular 3D voxel grids, voxel features of each non-empty regular 3D voxel grid are generated, and the discrete spatial index information of each non-empty regular 3D voxel grid is recorded to obtain voxel coordinates. In this invention, the size range of the voxel grid is (0.2m, 0.2m, 0.2m) to (0.4m, 0.4m, 0.4m). In one embodiment of this invention, the voxel size is 0.32m × 0.32m × 0.32m.

[0026] S12. Channel compression is performed on voxel features to obtain voxel descriptors, and a concise feature map is generated based on all voxel descriptors. More specifically, the method for channel compression of voxel features to obtain voxel descriptors includes sequential linear projection processing, max pooling processing, and layer normalization processing.

[0027] Furthermore, the specific content and effects of linear projection processing, max pooling processing, and layer normalization processing are as follows.

[0028] Linear projection processing uses a learnable linear transformation matrix to map the original high-dimensional voxel features to a low-dimensional space, thereby reorganizing and optimizing the voxel features and enabling information interaction and fusion between different feature channels. The parameters of linear projection are automatically learned through an end-to-end training process. The specific implementation method of linear projection processing is existing technology and will not be elaborated here.

[0029] Max pooling takes the maximum value of each channel at all points within each regular 3D voxel grid to preserve the target signal and suppress noise, while mitigating the impact of point position errors. After max pooling, each voxel is compressed from a set that may contain multiple points into a fixed-dimensional voxel descriptor. The specific timing method of max pooling is existing technology and will not be elaborated here.

[0030] Layer normalization standardizes the compressed voxel descriptors to make their mean 0 and variance 1, thereby stabilizing the training process, accelerating model convergence, and enhancing the model's adaptability to features of different dimensions. The specific implementation of layer normalization is based on existing technology and will not be elaborated here.

[0031] After processing by S12, the point cloud information that was originally scattered in the non-empty regular 3D voxel mesh is aggregated into a compact voxel descriptor, which preserves the local geometry of the target while reducing the data dimensionality.

[0032] S13. Reconstruct the simplified feature map based on voxel coordinates to obtain the optimized feature map. More specifically, when reconstructing the simplified feature map based on voxel coordinates, the PillarScatter3D module is used to reconstruct the optimized feature map. The specific processing method of the PillarScatter3D module is as follows.

[0033] First, voxel features across all height layers within each cylinder are maintained as independent channels, rather than aggregated into a single value. If the original regular 3D voxel mesh has H layers in the height direction, the generated BEV feature map will have H×C channels, where C is the number of channels for each voxel descriptor. Second, utilizing the positional information provided by the voxel coordinate branch, feature points from non-empty regular 3D voxel meshes are mapped to their corresponding BEV mesh positions, avoiding invalid calculations for empty voxels. The computational complexity is proportional to the number of points in the sparse point cloud of the scene. Finally, based on the discrete indices recorded by the voxel coordinate branch, the voxel features are mapped to their corresponding meshes on the BEV plane according to their coordinates.

[0034] After processing by S13, the generated high-aware BEV feature map has a dimensional structure of (H×C,W,H_bev), where W and H_bev are the width and height of the BEV grid, respectively.

[0035] S2. Using a pre-trained state-space model, the optimized feature map is processed in several stages to obtain several basic feature maps. All basic feature maps are fused to obtain a multi-scale fused feature map. The multi-scale fused feature map is then used to detect the target detection result.

[0036] In this invention, the state-space model adopts the Mamba2 model, which implements... The linear inference time complexity, compared to the traditional Transformer model. Compared to the complexity of Mamba2, this significantly reduces computational resource consumption, making it extremely suitable for processing long sequence features on resource-constrained edge devices. In this invention, the state-space model, based on Mamba2, introduces a selective state-space mechanism, transforming the sequence modeling problem into a recursive update of the state space. The state at each time step depends only on the state of the previous step and the current input, thus achieving… The linear complexity is achieved. To process the 2D BEV feature map, this embodiment first unfolds the 2D BEV feature map into a 1D token sequence at the beginning of each stage. The unfolding is done in row-major order, that is, each cell in the BEV grid is converted into a token in the sequence row by row, and each token corresponds to a feature vector of a BEV grid cell. In this way, the adjacency relationship in the 2D space is transformed into the proximity relationship in the 1D sequence, enabling the state-space model to utilize its state-space mechanism to model long-distance dependencies along the sequence direction.

[0037] Furthermore, the processing methods for the state-space model in each stage include S201 to S203.

[0038] S201. The input data is flattened to obtain a one-dimensional token sequence, which is then fed into the first and second processing paths in parallel. The input data is either an optimized feature map or a basic feature map output from the previous stage. The flattened one-dimensional token sequence is copied twice and fed into the first and second processing paths respectively. The two paths execute in parallel without interference, allowing the model to model the features from two perspectives simultaneously.

[0039] Furthermore, the length N of the one-dimensional token sequence is equal to the height multiplied by the width of the BEV grid. If the BEV size is set to the common 200×200, then N=40000. Employing the linear complexity mechanism of the state-space model, only about 40,000 state updates are required, a computational complexity far lower than that of the Transformer. On edge devices with multi-core CPUs or GPUs, the first and second processing paths can be executed in full parallel on different computing units to reduce inference latency.

[0040] S202. In the first processing path, the channel dimension of the one-dimensional token sequence is increased through feature projection. In the second processing path, the channel dimension of the one-dimensional token sequence is reduced through feature compression. More specifically, if the number of channels in the one-dimensional token sequence is C, the first processing path first expands the number of channels to 2C through a linear projection layer, providing a richer representation space for subsequent state space modeling. After state space model processing, the number of channels is restored from 2C to C through another linear projection. The second processing path first compresses the number of channels from C to a smaller value through feature compression to extract the essential low-dimensional structure from the high-dimensional features.

[0041] S203. Back-projection and feature fusion are performed on the first processing result of the first processing path and the second processing result of the second processing path to obtain a basic feature map. More specifically, during back-projection, the number of feature channels after processing by the state-space model in the first processing path is 2C, which is compressed back to C through back-linear projection; the number of feature channels after processing by the state-space model in the second processing path is C_small, which is expanded back to C through back-linear projection. Both the first and second processing paths use learnable weight matrices for back-projection and do not share parameters. After back-projection, the first and second processing paths yield two feature sequences with the same dimension C, which are fused element-wise to integrate the information from the two paths without increasing the channel dimension. The fused result is then layer-normalized to stabilize the distribution range of feature values, prevent excessively large or small activation values ​​from causing gradient vanishing or exploding, and accelerate model convergence. Furthermore, residual connections are established with the input data to ensure that the original input information can be directly transmitted to the output, ensuring that fine local details are not lost and facilitating gradient backpropagation. The one-dimensional feature sequence after residual connection processing is reverse-mapped according to the order used in the previous expansion and reshaped into a two-dimensional BEV structure.

[0042] Furthermore, during the processing of the optimized feature map by the state-space model, downsampling convolution is performed on the input data before any stage. Additionally, at the end of each stage, upsampling is performed on the basic feature map obtained in the previous step to restore the spatial resolution reduced by the downsampling convolution. In one embodiment of the invention, upsampling is achieved using bilinear interpolation combined with 1×1 convolution. Specifically, the backbone network of the state-space model of the present invention comprises three consecutive progressive stages, with each stage gradually reducing the spatial resolution through downsampling convolution to achieve multi-scale receptive field expansion. The first stage processes the optimized feature map at the original resolution to preserve low-level spatial details; the second stage processes the downsampling feature map to extract mid-level local structures; and the third stage processes the further downsampling feature map to capture high-level global semantics.

[0043] Furthermore, when fusing all basic feature maps, channel stitching is performed on all basic feature maps to form a multi-scale fused feature map. The specific process is as follows.

[0044] First, the feature maps generated in the second and third stages are upsampled to the same spatial resolution as the first stage. Specifically, the feature map generated in the second stage is upsampled by a factor of 2, and the feature map generated in the third stage is upsampled by a factor of 4. Second, after size alignment, the three feature maps have the same height and width. If the number of channels in the first stage is C1, the number of channels in the second stage is C2, and the number of channels in the third stage is C3, then the number of channels in the stitched multi-scale fused feature map is C1 + C2 + C3. For example, in one embodiment of the present invention, C1 = 64, C2 = 128, C3 = 256, and the number of channels in the fused feature map is 448. In scenarios where strict control of computational latency is required, the number of channels is further compressed to reduce the computational overhead of the target detection head.

[0045] like Figure 4 As shown, the inference time decomposition of this invention has significant advantages compared to other benchmark models. The RTNH+ model, due to its complex 3D convolutional network and Transformer module, has the longest computation time in its backbone network, accounting for over 60% of the total inference time, resulting in an overall frame rate of less than 15 FPS. The RadarPillars model reduces computation to some extent through its pillar attention mechanism, but its PillarAttention still has limitations. The complexity of the Mamba2 backbone network implementation in this invention. With linear complexity, the overall computational cost is far lower than RadarPillars for a typical sequence length of N=40000.

[0046] More specifically, this invention employs a hybrid voxel cylinder encoder, processing only non-empty voxels, with this part taking approximately 12ms, lower than RTNH+'s 18ms and RadarPillars' 15ms; the PillarScatter3D module of this invention uses sparse point mapping, with this part taking approximately 5ms, far lower than the 22ms required for the 3D convolution transpose operation in RTNH+; the total time of the three stages of the Mamba2 backbone network of this invention is approximately 25ms, lower than RadarPillars' 28ms and significantly lower than RTNH+'s 45ms; the detection head of this invention adopts a lightweight design, and NMS accelerates convergence due to the high quality of the candidate boxes, with a total time of approximately 7ms.

[0047] In summary, the total inference time per frame of this invention is approximately 49ms, corresponding to approximately 20.4 FPS, which meets the real-time requirement of 15-20 FPS for autonomous driving.

[0048] As shown in Table 1, the detection accuracy of this invention is superior to the benchmark model under various weather conditions. In foggy conditions, this invention achieves an AP3D accuracy of 87.8%, which is 9.3 percentage points higher than RadarPillars' 78.5%. In rainy conditions, this invention improves upon RadarPillars by 8.1 percentage points. In heavy snow, this invention achieves an accuracy of 60.0%, while RTNH+ only achieves 44.5%. Experimental results show that the hybrid voxel cylinder encoder of this invention effectively suppresses clutter noise through max pooling, and the Mamba2 backbone can filter invalid sequence information, effectively extracting target features even in extremely noisy environments.

[0049] Table 1. Performance Comparison of the Invention and Prior Art

[0050] The present invention further provides a target detection system based on 4D LiDAR to implement the above-mentioned target detection method based on 4D LiDAR. The system includes a data acquisition module, a point cloud processing module and a target detection module.

[0051] The data acquisition module is used to acquire sparse point clouds from 4D LiDAR.

[0052] The point cloud processing module is used to process sparse point clouds using a hybrid voxel cylinder encoder to obtain optimized feature maps.

[0053] The object detection module is used to process the optimized feature map in several stages using a pre-trained state space model to obtain several basic feature maps, fuse all the basic feature maps to obtain a multi-scale fused feature map, and detect the multi-scale fused feature map to obtain the object detection result.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or modules may be electrical, mechanical, or other forms.

[0055] The present invention also provides a computer device, including a memory and a processor.

[0056] Memory is used to store computer programs.

[0057] A processor is used to read and execute computer programs to implement the aforementioned target detection method based on 4D LiDAR.

[0058] Finally, the present invention provides a storage medium for storing a computer program that, when executed, implements the above-described target detection method based on 4D LiDAR.

[0059] The memory, as a carrier of resources, can be a read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include operating systems, computer programs, etc., and the storage method can be temporary or permanent storage. The operating system is used to manage and control the various hardware devices and computer programs on the electronic device; it can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the methods disclosed in any of the foregoing embodiments, the computer programs may further include computer programs capable of performing other specific tasks. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A target detection method based on 4D lidar, characterized in that, include: The sparse point cloud output from the 4D LiDAR is acquired and input into a hybrid voxel cylinder encoder for processing. The processing method of the hybrid voxel cylinder encoder for the sparse point cloud includes: The sparse point cloud is voxelized to obtain the original feature map, and voxel features and voxel coordinates are generated. Voxel features are channel compressed to obtain voxel descriptors, and a concise feature map is generated based on all voxel descriptors; An optimized feature map is obtained by reconstructing a simplified feature map based on voxel coordinates; The process involves using a pre-trained state-space model to process the optimized feature map in several stages to obtain several basic feature maps, fusing all the basic feature maps to obtain a multi-scale fused feature map, and then performing detection on the multi-scale fused feature map to obtain the target detection result. The state-space model processing methods in each stage include: The input data is flattened to obtain a one-dimensional token sequence and fed into the first and second processing paths in parallel. The input data is either an optimized feature map or a basic feature map output from the previous stage. In the first processing path, the channel dimension of the one-dimensional token sequence is increased by feature projection, and in the second processing path, the channel dimension of the one-dimensional token sequence is reduced by feature compression. The basic feature map is obtained by back-projecting and feature fusion of the first processing result of the first processing path and the second processing result of the second processing path.

2. The target detection method based on 4D lidar as described in claim 1, characterized in that, Methods for obtaining the original feature map by voxelization of sparse point clouds include: Regions of interest (ROIs) are obtained by cropping sparse point clouds. The region of interest is divided into multiple regular 3D voxel grids to obtain the original feature map.

3. The target detection method based on 4D lidar as described in claim 2, characterized in that, When the region of interest is divided into multiple regular 3D voxel grids, the voxel features of each non-empty regular 3D voxel grid are generated, and the discrete spatial index information of each non-empty regular 3D voxel grid is recorded to obtain the voxel coordinates.

4. The target detection method based on 4D lidar as described in claim 1, characterized in that, The method for obtaining voxel descriptors by channel compression of voxel features includes sequential linear projection processing, max pooling processing, and layer normalization processing.

5. The target detection method based on 4D lidar as described in claim 1, characterized in that, When reconstructing a simplified feature map based on voxel coordinates, the PillarScatter3D module is used to reconstruct and obtain an optimized feature map.

6. The target detection method based on 4D lidar as described in claim 1, characterized in that, In the process of optimizing feature maps, the state-space model performs downsampling convolution on the input data before any stage.

7. The target detection method based on 4D lidar as described in claim 1, characterized in that, When fusing all basic feature maps, channel stitching is performed on all basic feature maps to form a multi-scale fused feature map.

8. A target detection system based on 4D lidar, characterized in that, For implementing a target detection method based on 4D lidar as described in any one of claims 1-7, the system comprises: The data acquisition module is used to acquire sparse point clouds from 4D LiDAR; The point cloud processing module is used to process sparse point clouds using a hybrid voxel cylinder encoder to obtain optimized feature maps. The object detection module is used to process the optimized feature map in several stages using a pre-trained state space model to obtain several basic feature maps, fuse all the basic feature maps to obtain a multi-scale fused feature map, and detect the multi-scale fused feature map to obtain the object detection result.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for reading and executing the computer program to implement a target detection method based on 4D LiDAR as described in any one of claims 1-7.

10. A storage medium, characterized in that, Used to store a computer program, which, when executed, implements a target detection method based on 4D LiDAR as described in any one of claims 1-7.

Images