A three-dimensional target detection method, system, device and storage medium
Patent Information
- Application Number
- CN202410236440.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-01
- Publication Date
- 2026-10-16
AI Technical Summary
[0010]为了解决现有三维三维目标检测方法检测精度低的问题，本发明提供了一种三维目标检测方法、系统、设备及存储介质
[0038]本发明提供的一种三维目标检测方法、系统、设备及存储介质具有以下有益效果：本发明提供的三维目标检测方法是融合了几何及语义信息的深度补全三维稀疏卷积目标检测方法，其能够应用于来自多传感器的多模态信息进行数据级融合以及特征级融合的感知算法，能够充分适应并利用不同模态数据的各自特性，提高检测性能，得到更加准确、可靠的目标检测信息，提高检测信息的可用性。
Smart Images
Figure CN122929548A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of three-dimensional target detection technology, specifically relating to a three-dimensional target detection method, system, device, and storage medium. Background Technology
[0002] 3D object detection is a key branch of computer vision, widely used in intelligent transportation, robotics, 3D modeling, virtual reality, and augmented reality. Its core objective is to determine the type, location, depth, size, and orientation of objects in 3D space. 3D object detection can be mainly divided into three categories: 2D image-based detection, point cloud-based detection, and multimodal detection.
[0003] In 3D object detection based on 2D images, points in 3D space are mapped onto a 2D plane, resulting in the loss of crucial depth information. Due to the lack of depth data, 2D images alone cannot accurately capture the full range of attributes of objects in 3D space, such as category, location, depth, size, and pose. This is a major difference between 3D and 2D object detection and one of the reasons why 3D object detection is more challenging.
[0004] 3D object detection based on 2D images attempts to detect 3D objects by directly extracting depth information from images. This collection of depth information relies heavily on geometric constraints of the 3D scene, constraints on object shape, and semantic constraints. However, because the depth information in 2D images is very limited, and these constraints can be significantly restricted in different scenes and with different objects, this type of 3D object detection method typically has low accuracy. These factors collectively highlight the complexity and limitations of 2D image-based 3D object detection.
[0005] A point cloud is a collection of points in three-dimensional space, corresponding to pixels in a two-dimensional image. 3D object detection based on point cloud data can acquire depth information and falls into two main categories:
[0006] The first category directly processes point cloud data in three-dimensional space. This involves extending pixel-level operations from two-dimensional object detection to three dimensions, enabling direct manipulation of the point cloud. This method has higher computational complexity due to the increased dimensionality of the computation. Furthermore, noisy data in the point cloud can directly affect detection accuracy.
[0007] The second approach first trains a depth prediction model using point cloud data, and then uses this model to obtain two-dimensional depth images. Depth information is then extracted from these depth images for three-dimensional object detection. This method does not directly manipulate point cloud data, but rather reduces it to a two-dimensional depth map, thus reducing computational complexity. The depth prediction model can also eliminate some point cloud noise, making it more commonly used in practical applications.
[0008] One existing 3D object detection technique, after obtaining a 2D depth image, uses a depth prediction model trained on point cloud data, and then further trains a 3D object detection model using 2D RGB images. The limitation of this method is that, for 3D object detection, directly processing 3D point cloud data to obtain object category and location information from 2D images or video frames is unnecessary, and point cloud data typically contains a large amount of noise.
[0009] Another existing technique is to use a 2D depth image as a separate model input, extract depth information from the depth image using an additional model, and combine it with a 2D RGB (red, green, blue) image input for 3D object detection. The drawback of this method is that the depth information obtained from the 2D image is very limited, and geometric constraints are inevitably involved in the process of acquiring depth information, thus reducing detection accuracy. Summary of the Invention
[0010] To address the problem of low detection accuracy in existing 3D target detection methods, this invention provides a 3D target detection method, system, device, and storage medium.
[0011] To achieve the above objectives, the present invention provides the following technical solution:
[0012] A three-dimensional target detection method includes the following steps:
[0013] Obtain the raw point cloud data and RGB image data of the target image;
[0014] The points in the original point cloud data are augmented and completed using RGB image data to obtain semantically enhanced point cloud data. The original point cloud data is voxelized to obtain voxelized point cloud data.
[0015] The semantically enhanced point cloud data is used to perform a three-dimensional sparse convolution operation to obtain multi-scale three-dimensional sparse point cloud features containing spatial feature information. The voxelized point cloud data is used to perform a three-dimensional sparse convolution operation to obtain voxel point cloud features containing both semantic and spatial information.
[0016] By fusing multi-scale 3D sparse point cloud features containing spatial information with voxel point cloud features containing semantic and spatial information, complete 3D target detection information is obtained.
[0017] Furthermore, the target image includes pedestrians, motor vehicles, and non-motor vehicles around the car.
[0018] Furthermore, before using RGB image data to perform data augmentation and completion on the points in the point cloud to obtain semantically enhanced point cloud data, and before voxelizing the point cloud data to obtain voxelized point cloud data, data processing is required on the point cloud data and RGB image data. The data processing includes normalizing the coordinates of the image and labeling the categories of objects in the image and the coordinates of the two-dimensional detection box of the image, normalizing the coordinates of the point cloud and labeling the categories of objects in the point cloud, as well as the center position, size, and orientation angle of the three-dimensional detection box.
[0019] Furthermore, the following methods are used to augment and complete the points in the point cloud using RGB image data to obtain semantically enhanced point cloud data:
[0020] Point cloud data and RGB image data are input into a depth completion network. Data augmentation and completion are performed on points P = {p}, p = {x, y, z, α} in the point cloud, and semantically augmented point cloud is output. The data format of each point in the semantically augmented point cloud is as follows:
[0021] p*=(x,y,z,u,v,r,g,b)
[0022] Where x, y, z are the coordinates of the points in the point cloud data, u, v are the corresponding two-dimensional coordinates of the points in the image data, used to measure the relationship between the point and nearby points in two dimensions, and r, g, b are the color information of the points in the image data.
[0023] Furthermore, the multi-scale 3D sparse point cloud features containing spatial characteristic information are obtained by performing a 3D sparse convolution operation on the semantically enhanced point cloud data. The 3D sparse convolution operation is implemented by the following formula:
[0024] F s =SpConv s (V;W s )
[0025] Among them: W s is the convolution weight of the s-th layer, and V represents the voxel size;
[0026] The multi-scale 3D sparse point cloud features are achieved by connecting the 3D sparse features at each scale using the following formula:
[0027] F conca =Concat(F1,F2,…,F s )t
[0028] Wherein: F s These are features obtained at the scale of the s-th layer.
[0029] Furthermore, the three-dimensional sparse features at each scale are normalized to obtain the region of interest (ROI), and the semantically enhanced point cloud data and voxelized point cloud data are segmented within the ROI.
[0030] Furthermore, the complete 3D target detection information includes the category, position, and pose information of the 3D objects in the target image.
[0031] A target detection system, comprising:
[0032] The data acquisition module is used to acquire the raw point cloud data and RGB image data of the target image;
[0033] The enhancement and completion module is used to enhance and complete the points in the original point cloud data using RGB image data to obtain semantically enhanced point cloud data, and to voxelize the original point cloud data to obtain voxelized point cloud data.
[0034] The feature extraction module is used to perform a three-dimensional sparse convolution operation on the semantically enhanced point cloud data to obtain multi-scale three-dimensional sparse point cloud features containing spatial feature information, and to perform a three-dimensional sparse convolution operation on the voxelized point cloud data to obtain voxel point cloud features containing semantic information and spatial information.
[0035] The feature fusion detection module is used to fuse multi-scale 3D sparse point cloud features containing spatial feature information and voxel point cloud features containing semantic and spatial information to obtain complete 3D target detection information.
[0036] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the computer device performs the steps of the three-dimensional target detection method as described in any of the preceding claims.
[0037] A storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the three-dimensional target detection method as described in any of the preceding claims.
[0038] The three-dimensional target detection method, system, device, and storage medium provided by this invention have the following beneficial effects: The three-dimensional target detection method provided by this invention is a depth-complete three-dimensional sparse convolutional target detection method that integrates geometric and semantic information. It can be applied to perception algorithms that perform data-level fusion and feature-level fusion of multi-modal information from multiple sensors. It can fully adapt to and utilize the characteristics of different modal data, improve detection performance, obtain more accurate and reliable target detection information, and improve the availability of detection information. Attached Figure Description
[0039] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 This is a flowchart of a three-dimensional target detection method according to an embodiment of the present invention;
[0041] Figure 2 This is a flowchart illustrating the parameters related to training the network model of this invention;
[0042] Figure 3 This is a diagram of the backbone network structure of the three-dimensional target detection method according to an embodiment of the present invention. Detailed Implementation
[0043] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0044] This invention provides a depth-complete 3D sparse convolutional target detection method that integrates geometric and semantic information. It can be applied to perception algorithms that perform data-level fusion and feature-level fusion of multimodal information from multiple sensors. It can fully adapt to and utilize the characteristics of different modal data, improve detection performance, obtain more accurate and reliable target detection information, and improve the safety of intelligent vehicle systems.
[0045] Method Implementation Examples:
[0046] like Figure 1 , 2 As shown, a three-dimensional target detection method of the present invention includes the following steps:
[0047] S1. Obtain point cloud data and RGB image data of the target image, where the target image includes pedestrians, motor vehicles, and non-motor vehicles around the car. Construct training and testing set data using the RGB images and point cloud data, including:
[0048] RGB images and point cloud data are acquired and divided into training and testing sets at an approximately 1:1 ratio. Based on the image data in the training and testing sets, the coordinates of the images are normalized, and the categories of objects in the images and the coordinates of the 2D detection boxes are labeled. The coordinates of the point clouds are normalized, and the categories of objects in the point clouds, as well as the center position, size, and orientation angle of the 3D detection boxes, are labeled.
[0049] S2. Use RGB image data to perform data augmentation and completion on the points in the point cloud to obtain semantically enhanced point cloud data. Perform voxelization on the point cloud data to obtain voxelized point cloud data. The steps are as follows: Input the point cloud data and RGB image data into the trained deep completion network, perform data augmentation and completion on the points in the point cloud to obtain semantically enhanced point cloud data.
[0050] The point cloud data and RGB image data from the training and testing sets are input into the trained deep completion network. Data augmentation and completion are performed on the points P = {p}, p = {x, y, z, α} in the point cloud, and the semantically augmented point cloud is output. The data format of each point in the semantically augmented point cloud is as follows:
[0051] p* = (x, y, z, u, v, r, g, b),
[0052] Where x, y, z are the coordinates of the points in the point cloud data, u, v are the corresponding two-dimensional coordinates of the points in the image data, used to measure the relationship between the point and nearby points in two dimensions, and r, g, b are the color information of the points in the image data.
[0053] In addition, before using RGB image data to perform data augmentation and completion on points in the point cloud to obtain semantically enhanced point cloud data, and before voxelizing the point cloud data to obtain voxelized point cloud data, data processing is required on the point cloud data and RGB image data. Data processing includes normalizing the coordinates of the image and labeling the categories of objects in the image and the coordinates of the two-dimensional detection boxes in the image, normalizing the coordinates of the point cloud and labeling the categories of objects in the point cloud, as well as the center position, size and orientation angle of the three-dimensional detection boxes.
[0054] S3. Obtain multi-scale 3D sparse point cloud features containing spatial feature information using semantically enhanced point cloud data, and obtain voxel point cloud features containing semantic and spatial information using voxelized point cloud data.
[0055] S4. The multi-scale 3D sparse point cloud features containing spatial feature information and the voxel point cloud features containing semantic and spatial information are fused to obtain complete 3D target detection information. The complete 3D target detection information includes the category, position and pose information of the 3D object in the target image.
[0056] Specific examples Figure 3 As shown: In a 3D object detection model that includes a semantically enhanced point cloud backbone network, a voxel point cloud backbone network, an attention-based cross-membrane feature fusion processor, a classifier, and a regressor:
[0057] The main function of the voxel point cloud backbone network is to extract features from voxelized point clouds. This process involves performing multiple, multi-level 3D sparse convolution operations on the voxelized point cloud data, and then concatenating and normalizing the 3D sparse features obtained at each scale to obtain multi-scale 3D sparse point cloud features containing spatial feature information. Specifically:
[0058] Suppose the original point cloud consists of P = {p1, p2, ..., p...} N} represents each point p i ∈R 4 Voxelization allows a continuous spatial dimension to be discretized into a voxel mesh. Let the voxel size be V, then the voxelized point cloud can be represented as:
[0059]
[0060] 3D sparse convolution operation: For each scale s, the feature map can be obtained through 3D sparse convolution:
[0061] F s =SpConv s (V;W s )
[0062] Among them W s is the convolution weight of the s-th layer.
[0063] Multi-scale connectivity: Connecting 3D sparse features at different scales yields:
[0064] F concat =Concat(F1,F2,…,F s )
[0065] Wherein: F s These are features obtained at the scale of the s-th layer.
[0066] Normalization operation: The features are normalized using a normalization layer. Batch Normalization is then applied to obtain the features of the voxelized point cloud.
[0067] F voxel =BN(F concat ).
[0068] Based on the obtained multi-scale 3D sparse point cloud features, after dimensionality processing, non-maximum suppression (NMS), and alignment and adjustment, the region of interest (ROI) is obtained.
[0069] Based on the Region of Interest (ROI), the voxelized point cloud and semantically enhanced point cloud are segmented, with a focus on more refined detection, recognition, and classification of objects within the ROI.
[0070] Furthermore, the semantically enhanced point cloud backbone network extracts features from the semantically enhanced point cloud based on the region of interest, obtaining features containing semantic and spatial information. In this process, feature extraction is performed on points in the enhanced point cloud with color information. Specific processing is applied to the input data p*=(x,y,z,α,r,g,b), calculating differences from neighboring points, and performing specific indexing and shape operations to convert the input into a given number of channels. Then, a point convolutional network is used to extract features. The feature extraction process of the point convolutional network is as follows: first, the convolutional network extracts the enhanced point cloud features, then weights are calculated: weights are calculated using the convolutional network and neighboring points, then the features and weights are multiplied, and then the convolutional network is used to fuse them, concatenating the outputs of all layers to form the final feature vector.
[0071] Furthermore, a cross-modal feature fusion processor based on an attention mechanism performs high-precision feature fusion between features extracted from the voxel point cloud backbone network and features extracted from the semantically enhanced point cloud backbone network. Simultaneously, it regresses target category and 3D location information to obtain complete 3D target detection information, including:
[0072] A multimodal data fusion network based on a 3D mesh is constructed, which fuses the features extracted by the voxel point cloud backbone network and the features extracted by the semantically enhanced point cloud backbone network using the attention mechanism of the transformer.
[0073] The obtained feature fusion information is subjected to fusion feature extraction, and after unifying the scale, all feature information is connected together, and the attention mechanism is used to realize the corresponding information weight allocation.
[0074] By using a 3D object detection head to regress object category and pose information, calculating the corresponding loss function, training and optimizing the network, complete 3D object detection information is obtained.
[0075] Further explanation of the 3D target detection model:
[0076] 1. The semantically enhanced point cloud backbone network, voxel point cloud backbone network, attention-based cross-modal feature fusion, classifier and regressor in the three-dimensional object detection model all contain convolutional layers, fully connected layers and normalization layers. The semantically enhanced point cloud backbone network contains 4 semantic virtual sparse convolutional modules, and the voxel point cloud backbone network contains 4 virtual sparse convolutional modules.
[0077] 2. Train a 3D object detection model using training and testing sets. Semantic-enhanced point cloud backbone networks and voxel point cloud backbone networks acquire semantic-enhanced 3D features and voxel point cloud features, respectively. Input these features into a cross-modal self-attention learning module to obtain novel fused 3D features. Use these fused 3D features to learn a classifier and regressor, resulting in a trained 3D object detection model. The specific steps are as follows:
[0078] Initialize the parameters of the semantically enhanced point cloud backbone network, voxel point cloud backbone network, and the convolutional, fully connected, and normalized layers contained in the classifier and regressor in the 3D object detection model.
[0079] Set the relevant training parameters for the stochastic gradient descent algorithm, including the learning rate, impulse, batch size, and number of iterations.
[0080] For any given batch of iterations, all RGB images and point clouds are input into the data fusion engine to obtain semantically enhanced point clouds and voxelized point clouds. These are then input into the semantically enhanced point cloud backbone network and the voxelized point cloud backbone network, respectively, to obtain semantically enhanced point cloud features and voxelized point cloud features. A cross-modal self-attention learning module is constructed. The semantically enhanced point cloud features and voxelized point cloud features are then input into the cross-modal self-attention feature fusion module to obtain fused features. The RGB features are updated using the self-attention matrix. The updated classifier and regressor are then used for target detection of 3D objects based on multimodal data. The objective function value is obtained by calculating the error between the network estimate and the actual labeled value. The objective function value is then calculated using a loss function. The loss function is as follows:
[0081] L = L rpn +L roi +λ1L vox +λ2L enh +βL depth
[0082] In this formula, λ1, λ2, and β are respectively assigned to L vox L enh and L depth The weights are set to default values of λ1 = 0.5, λ2 = 0.5, and β = 1, respectively. The losses of the RPN and RoI heads are expressed as L... rpn and L roi The depth completion loss is represented as L depth To ensure unbiased training across different processes, we introduce additional RoI head losses for both the voxel point cloud flow and the augmented point cloud flow, denoted as Li. vox and L enh It is worth noting that L vox and L enhIn terms of construction, it is similar to L roi Similarities exist, including both the category confidence loss and their respective regression loss components.
[0083] System Implementation Example:
[0084] This invention also provides a target detection system suitable for the above-described three-dimensional target detection method, comprising:
[0085] Data acquisition module: Used to acquire target image data.
[0086] Data fusion module: used to obtain semantically enhanced point cloud data and voxelized point cloud data from target image data.
[0087] Semantic Enhancement Point Cloud Backbone Network Feature Extraction Module: Used to obtain multi-scale 3D sparse point cloud features containing spatial feature information using semantically enhanced point cloud data.
[0088] Voxel point cloud backbone network feature extraction module: used to obtain voxel point cloud features containing semantic and spatial information using voxelized point cloud data.
[0089] Feature fusion detection module: This module is used to fuse multi-scale 3D sparse point cloud features containing spatial feature information and voxel point cloud features containing semantic and spatial information to obtain complete 3D target detection information.
[0090] Computer equipment example:
[0091] A computer device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it causes the computer device to perform the steps of a three-dimensional target detection method. A three-dimensional target detection method has been described sufficiently in the method embodiments and will not be repeated here.
[0092] Storage medium examples:
[0093] A storage medium storing a computer program, which, when executed by a processor, implements the steps of a three-dimensional target detection method. The three-dimensional target detection method has been described sufficiently in the method embodiments and will not be repeated here.
[0094] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0095] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0098] It should be noted that the above specific embodiments enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention, and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A three-dimensional target detection method, characterized in that, Includes the following steps: Obtain the raw point cloud data and RGB image data of the target image; The points in the original point cloud data are augmented and completed using RGB image data to obtain semantically enhanced point cloud data. The original point cloud data is voxelized to obtain voxelized point cloud data. The semantically enhanced point cloud data is used to perform a three-dimensional sparse convolution operation to obtain multi-scale three-dimensional sparse point cloud features containing spatial feature information. The voxelized point cloud data is used to perform a three-dimensional sparse convolution operation to obtain voxel point cloud features containing both semantic and spatial information. By fusing multi-scale 3D sparse point cloud features containing spatial information with voxel point cloud features containing semantic and spatial information, complete 3D target detection information is obtained.
2. The three-dimensional target detection method according to claim 1, characterized in that, The target image includes pedestrians, motor vehicles, and non-motor vehicles around the car.
3. The three-dimensional target detection method according to claim 1, characterized in that, Before using RGB image data to perform data augmentation and completion on points in the original point cloud data to obtain semantically enhanced point cloud data, and before voxelizing the point cloud data to obtain voxelized point cloud data, data processing is required on the point cloud data and RGB image data. The data processing includes normalizing the coordinates of the image and labeling the categories of objects in the image and the coordinates of the two-dimensional detection box of the image, normalizing the coordinates of the point cloud and labeling the categories of objects in the point cloud, as well as the center position, size and orientation angle of the three-dimensional detection box.
4. The three-dimensional target detection method according to claim 3, characterized in that, Using RGB image data to perform data augmentation and completion on points in a point cloud to obtain semantically enhanced point cloud data, specifically: Point cloud data and RGB image data are input into a depth completion network. Data augmentation and completion are performed on points P = {p}, p = {x, y, z, α} in the point cloud, and semantically augmented point cloud is output. The data format of each point in the semantically augmented point cloud is as follows: p*=(x,y,z,u,v,r,g,b) Where x, y, z are the coordinates of the points in the point cloud data, u, v are the corresponding two-dimensional coordinates of the points in the image data, used to measure the relationship between the point and nearby points in two dimensions, and r, g, b are the color information of the points in the image data.
5. The three-dimensional target detection method according to claim 4, characterized in that, The three-dimensional sparse convolution operation is implemented by the following formula: F s =SpConv s (V;W s ) Among them: W s is the convolution weight of the s-th layer, and V represents the voxel size; The multi-scale 3D sparse point cloud features are achieved by connecting the 3D sparse features at each scale using the following formula: Wherein: F s These are features obtained at the scale of the s-th layer.
6. The three-dimensional target detection method according to claim 5, characterized in that, The three-dimensional sparse features at each scale are normalized to obtain the region of interest (ROI). The semantically enhanced point cloud data is then segmented within the ROI. After segmentation, the three-dimensional sparse features at each scale are extracted to obtain the three-dimensional sparse features at each scale.
7. The three-dimensional target detection method according to claim 1, characterized in that, The 3D target detection information includes the category, position, and pose information of the 3D objects in the target image.
8. A target detection system, characterized in that, include: The data acquisition module is used to acquire the raw point cloud data and RGB image data of the target image; The enhancement and completion module is used to enhance and complete the points in the original point cloud data using RGB image data to obtain semantically enhanced point cloud data, and to voxelize the original point cloud data to obtain voxelized point cloud data. The feature extraction module is used to perform a three-dimensional sparse convolution operation on the semantically enhanced point cloud data to obtain multi-scale three-dimensional sparse point cloud features containing spatial feature information, and to perform a three-dimensional sparse convolution operation on the voxelized point cloud data to obtain voxel point cloud features containing semantic information and spatial information. The feature fusion detection module is used to fuse multi-scale 3D sparse point cloud features containing spatial feature information and voxel point cloud features containing semantic and spatial information to obtain complete 3D target detection information.
9. A computer device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it causes the computer device to perform the steps of the three-dimensional target detection method as described in any one of claims 1-7.
10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the three-dimensional target detection method as described in any one of claims 1-7.