Three-dimensional target detection method and device

Abstract

The invention relates to the technical field of computer vision, in particular to a three-dimensional target detection method and device, computer equipment, a storage medium and a computer program product. Acquiring point cloud data of a to-be-detected target in a to-be-detected space, and performing voxelization processing on the point cloud data to obtain voxel point cloud data; performing feature extraction on the voxel point cloud data by using a pre-constructed first feature network to obtain a three-dimensional positioning feature and a target recognition feature corresponding to the voxel point cloud data; projecting the to-be-measured target into a preset two-dimensional plane based on the three-dimensional positioning features, and obtaining two-dimensional positioning features of the to-be-measured target; obtaining positioning parameters of a three-dimensional detection frame corresponding to the to-be-detected target based on the two-dimensional positioning features; and optimizing the three-dimensional detection frame based on the target identification feature to obtain a target detection result. By adopting the method, the precision and the detection efficiency of target detection in a three-dimensional space can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a three-dimensional target detection method, apparatus, computer equipment, storage medium and computer program product. Background Art

[0002] Autonomous driving refers to the use of artificial intelligence, visual computing, radar, monitoring devices and global positioning systems to enable computers to automatically and safely operate motor vehicles without any active human operation. In autonomous driving, the detection of 3D targets in the environment is a very important technical module. This technical module is based on computer vision technology and is also the core basic direction of computer vision research, so it has received widespread attention. Unlike image recognition, target detection not only needs to identify the objects in the image and give the corresponding categories, but also needs to give the location of the object in the form of a minimum bounding box. Depending on the output results required by target detection, RGB images are generally used for target detection, and the method of outputting object categories and the minimum bounding box on the image is called 2D target detection, while the detection of RGB images, RGB-D depth images and laser point clouds, outputting object categories and information such as length, width, height, and rotation angle in three-dimensional space is called 3D target detection. For example, in the scenario of autonomous driving, it is necessary to provide indicators such as the three-dimensional size and rotation angle of the target object from the image, and the information in the bird's-eye view projection plays a vital role in path planning and control in subsequent autonomous driving scenarios.

[0003] In related technologies, with the widespread application of LiDAR sensors and the like, 3D scene information is captured as sparse and irregular point clouds, providing important 3D scene perception clues. Among them, in the common PVRCNN method, boxes with different sizes and aspect ratios set in advance by manual or clustering methods cover the entire image to prevent missed detection; the CenterPoint method is based on the Center-based two-stage detection combined with tracking model CenterPoint. In the first stage, a key point detector is used to detect the center of the object, and the 3D size, 3D direction, and speed of the detection box are regressed. In the second stage, a memory refinement module is designed to refine the detection box generated in the first stage using additional point features.

[0004] However, the current target detection methods have the following technical problems:

[0005] The PVRCNN method has problems of inaccurate center detection and low computational efficiency, while the CenterPoint method cannot adapt to the computing scenarios of sparse point clouds. It is difficult to achieve efficient and accurate target detection based on existing detection methods. Summary of the invention

[0006] Based on this, it is necessary to provide a three-dimensional target detection method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy and detection efficiency of target detection in three-dimensional space in response to the above technical problems.

[0007] In a first aspect, the present application provides a three-dimensional target detection method. The method comprises:

[0008] Acquire point cloud data of the target to be detected in the detection space, and voxelize the point cloud data to obtain voxel point cloud data;

[0009] Using a pre-constructed first feature network to perform feature extraction on the voxel point cloud data, and obtaining three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data;

[0010] Based on the three-dimensional positioning feature, the target to be measured is projected into a preset two-dimensional plane to obtain the two-dimensional positioning feature of the target to be measured;

[0011] Obtaining positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features;

[0012] The three-dimensional detection frame is optimized based on the target recognition feature to obtain a target detection result.

[0013] In one embodiment, the first feature network is a sparse three-dimensional convolutional network, and the use of the pre-constructed first feature network to extract features from the voxel point cloud data to obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data includes:

[0014] Performing multiple downsampling feature extraction on the voxel point cloud data based on multiple convolutional layers in the first feature network to obtain the three-dimensional positioning feature;

[0015] Through a preset voxel feature capture network, feature extraction is performed on the voxel point cloud data in each convolution layer to obtain the target recognition feature.

[0016] In one embodiment, obtaining the positioning parameters of the three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features includes:

[0017] Acquire a central target area corresponding to the target to be measured according to the two-dimensional positioning feature;

[0018] A distance parameter between the central target area and the edge of the target to be measured is determined, and the positioning parameter is obtained according to the central target area and the distance parameter.

[0019] In one embodiment, determining a distance parameter between the central target area and the edge of the target to be measured, and obtaining the positioning parameter according to the central target area and the distance parameter comprises:

[0020] Generate several groups of candidate positioning parameters according to the central target area and the corresponding distance parameters;

[0021] The loss values ​​between the candidate positioning parameters and the preset reference true values ​​are respectively calculated, and the positioning parameters are determined according to the loss values.

[0022] In one of the embodiments, the positioning parameters include the category, center position parameter, size parameter and direction parameter of the three-dimensional detection box.

[0023] In one of the embodiments, obtaining the positioning parameters of the three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features further includes:

[0024] The positioning parameters are obtained through a preset detection head network, and the detection head network is an anchor-free frame network based on a center point.

[0025] In a second aspect, the present application also provides a three-dimensional object detection device. The device comprises:

[0026] A voxel processing module is used to obtain point cloud data of the target to be detected in the detection space, and voxelize the point cloud data to obtain voxel point cloud data;

[0027] A feature extraction module, used to extract features from the voxel point cloud data using a pre-built first feature network, and obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data;

[0028] A feature projection module, used to project the target to be measured into a preset two-dimensional plane based on the three-dimensional positioning feature, so as to obtain the two-dimensional positioning feature of the target to be measured;

[0029] A detection and positioning module, used to obtain positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features;

[0030] The target optimization module is used to optimize the three-dimensional detection frame based on the target recognition feature to obtain a target detection result.

[0031] In one embodiment, the first feature network is a sparse three-dimensional convolutional network, and the feature extraction module includes:

[0032] A multi-level downsampling module, used for performing multiple downsampling feature extraction on the voxel point cloud data based on multiple convolutional layers in the first feature network to obtain the three-dimensional positioning feature;

[0033] The voxel feature module is used to extract features of the voxel point cloud data in each convolution layer through a preset voxel feature capture network to obtain the target recognition features.

[0034] In one embodiment, the detection and positioning module includes:

[0035] A center detection module, used for acquiring a center target area corresponding to the target to be detected according to the two-dimensional positioning feature;

[0036] The boundary distance module is used to determine the distance parameter between the central target area and the edge of the target to be measured, and obtain the positioning parameter according to the central target area and the distance parameter.

[0037] In one embodiment, the boundary distance module includes:

[0038] A candidate parameter module, used to generate several groups of candidate positioning parameters according to the central target area and the corresponding distance parameters;

[0039] The screening and determination module is used to respectively calculate the loss values ​​between the candidate positioning parameters and the preset reference true values, and determine the positioning parameters according to the loss values.

[0040] In one of the embodiments, the positioning parameters include the category, center position parameter, size parameter and direction parameter of the three-dimensional detection box.

[0041] In one embodiment, the detection and positioning module further includes:

[0042] The detection head module is used to obtain the positioning parameters through a preset detection head network, and the detection head network is an anchor-free frame network based on a center point.

[0043] In a third aspect, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps in the three-dimensional object detection method described in any one of the embodiments in the first aspect are implemented.

[0044] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in a three-dimensional object detection method as described in any one of the embodiments in the first aspect are implemented.

[0045] In a fifth aspect, the present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps in the three-dimensional object detection method as described in any one of the embodiments in the first aspect are implemented.

[0046] The above-mentioned three-dimensional target detection method, device, computer equipment, storage medium and computer program product can achieve the beneficial effects of the technical problems in the corresponding background technology by deducing the technical features in the exclusive rights:

[0047] The point cloud data in the space to be tested is obtained, and the point cloud data is voxelized to obtain voxel point cloud data. Then, the voxel point cloud data is extracted using the pre-constructed first feature network to obtain the three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data, so as to obtain the position of the target to be detected and the basic information of the target. Subsequently, the two-dimensional position of the target to be detected in the projection plane is determined in the projection plane, and then the detection frame of the target is determined according to the two-dimensional position and the three-dimensional positioning features at this time. At this time, a preliminary target detection result is obtained. At this time, the target detection result is mainly reflected in the positioning result of the target position, that is, the positioning parameter of the three-dimensional detection frame. Finally, the target recognition features obtained in the first feature network are used to optimize the target detection results, enrich the content of the target detection results, and thus improve the detection effect of the target detection. In implementation, compared with the prior art, it can achieve the beneficial effect of improving the target detection effect and target detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] FIG1 is a schematic diagram of a first process of a three-dimensional object detection method in one embodiment;

[0049] Figure 2 is a second flow chart of a three-dimensional object detection method in another embodiment;

[0050] Figure 3 is a third flow chart of a three-dimensional object detection method in another embodiment;

[0051] Figure 4 is a schematic diagram of a fourth process of a three-dimensional object detection method in another embodiment;

[0052] Figure 5 is a fifth flow chart of a three-dimensional object detection method in another embodiment;

[0053] Figure 6 is a structural block diagram of a three-dimensional object detection device in one embodiment;

[0054] FIG. 7 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] In related technologies, with the widespread application of LiDAR sensors and the like, 3D scene information is captured as sparse and irregular point clouds, providing important 3D scene perception clues. Among them, in the common PVRCNN method, boxes with different sizes and aspect ratios set in advance by manual or clustering methods cover the entire image to prevent missed detection; the CenterPoint method is based on the Center-based two-stage detection combined with tracking model CenterPoint. In the first stage, a key point detector is used to detect the center of the object, and the 3D size, 3D direction, and speed of the detection box are regressed. In the second stage, a memory refinement module is designed to refine the detection box generated in the first stage using additional point features.

[0057] However, the current target detection methods have the following technical problems:

[0058] The PVRCNN method has the problems of inaccurate center detection and low computational efficiency, while the CenterPoint method cannot adapt to the computing scenarios of sparse point clouds. It is difficult to achieve efficient and accurate target detection based on existing detection methods. Specifically, PVRCNN is a new method for accurately detecting three-dimensional objects from point clouds. This method covers the entire image with boxes of different sizes and aspect ratios that are set manually or by clustering methods in advance to prevent missed detection. However, the target box generated by the anchor-based detection head takes a longer time to ensure a more rigorous screening step, and compared with the center-based detection mode, there is a problem of unstable center detection. When multi-scale 3D voxel CNN features are integrated using the voxel setabstraction layer, it often leads to greater positional offsets, especially for small targets (such as pedestrians and bicycles). The CenterPoint method uses a center-based, anchor-free detection head to identify targets. A large number of experiments have shown that the effect is not very good when the point cloud is too sparse. For example, the effect on the nuScenes dataset is not as good as Waymo. The main reason is that the center point detection method cannot effectively identify 3D point cloud data with too little information retained after feature extraction.

[0059] Based on this, the present application provides a three-dimensional target detection method.

[0060] In one embodiment, as shown in FIG1 , a three-dimensional object detection method is provided. This embodiment uses the method applied to a terminal as an example for illustration. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0061] Step 101: acquiring point cloud data of a target to be detected in a detection space, and performing voxel processing on the point cloud data to obtain voxel point cloud data.

[0062] The space to be detected may refer to the three-dimensional space area that can be covered by the sensing range for detecting the target. Point cloud data may refer to a set of vectors in a three-dimensional coordinate system, which is used to record the three-dimensional coordinates, color information, reflection intensity information and other sensing information of each point. Voxel may refer to a volume element, and a voxel may represent a three-dimensional area with a constant scalar or vector.

[0063] Exemplarily, the terminal can obtain the original point cloud data through a vehicle-mounted laser radar, etc., and then the terminal can transform the original point cloud data into point cloud data in a three-dimensional space that conforms to the vehicle coordinate system through coordinate transformation and other processing. Subsequently, the terminal can voxelize the point cloud data, thereby converting the sparse three-dimensional point cloud into a dense three-dimensional voxel feature, that is, voxel point cloud data. In this process, the terminal can assign each point cloud point in the point cloud data to a corresponding voxel. The number of point cloud points in each voxel can be one, multiple, or zero.

[0064] Step 103: Use the pre-constructed first feature network to perform feature extraction on the voxel point cloud data to obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data.

[0065] The first feature network may refer to a convolutional neural network (CNN) for extracting three-dimensional features. The three-dimensional positioning feature may refer to a parameter feature for determining the position and posture of a target. The target recognition feature may refer to a parameter for describing other basic information of the target object, such as color, material, etc.

[0066] Exemplarily, the terminal may utilize a pre-built first feature network to perform feature extraction on voxel point cloud data, thereby obtaining three-dimensional positioning features and target recognition features output by the model.

[0067] Step 105: projecting the target to be measured into a preset two-dimensional plane based on the three-dimensional positioning feature to obtain the two-dimensional positioning feature of the target to be measured.

[0068] The two-dimensional positioning feature may refer to the positioning feature of the target to be measured in a specific two-dimensional plane.

[0069] Exemplarily, the terminal may predetermine a general two-dimensional plane, which may be a bird's-eye view plane from the perspective of the vehicle body, or other two-dimensional planes. The terminal may project the target to be measured into the preset two-dimensional plane according to the three-dimensional positioning feature, thereby obtaining the two-dimensional positioning feature of the target to be measured. The two-dimensional positioning feature may be a heatmap.

[0070] Step 107: obtaining positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features.

[0071] The three-dimensional detection frame may refer to a bounding box of the target to be detected in the three-dimensional space.

[0072] For example, the terminal can determine the two-dimensional detection frame of the target in a preset two-dimensional plane according to the two-dimensional positioning feature, and then expand the determined two-dimensional detection frame according to the three-dimensional positioning feature to obtain the positioning parameters of the three-dimensional detection frame. The positioning parameters may include the category, center position parameter, size parameter and direction parameter of the three-dimensional detection frame.

[0073] Step 109: Optimize the three-dimensional detection frame based on the target recognition feature to obtain a target detection result.

[0074] Exemplarily, after obtaining the three-dimensional detection frame of the target to be detected, the terminal can use the target recognition features obtained in the previous step to perform rendering optimization on the current three-dimensional detection frame, thereby obtaining the final target detection result.

[0075] In the above three-dimensional target detection method, reasonable deduction combined with the technical features in the embodiment can achieve the following beneficial effects of solving the technical problems raised in the background technology:

[0076] The point cloud data in the space to be tested is obtained, and the point cloud data is voxelized to obtain voxel point cloud data. Then, the voxel point cloud data is extracted using the pre-constructed first feature network to obtain the three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data, so as to obtain the position of the target to be detected and the basic information of the target. Subsequently, the two-dimensional position of the target to be detected in the projection plane is determined in the projection plane, and then the detection frame of the target is determined according to the two-dimensional position and the three-dimensional positioning features at this time. At this time, a preliminary target detection result is obtained. At this time, the target detection result is mainly reflected in the positioning result of the target position, that is, the positioning parameter of the three-dimensional detection frame. Finally, the target recognition features obtained in the first feature network are used to optimize the target detection results, enrich the content of the target detection results, and thus improve the detection effect of the target detection. In implementation, compared with the prior art, it can achieve the beneficial effect of improving the target detection effect and target detection efficiency.

[0077] In one embodiment, Figure 2 As shown, the first feature network may be a sparse three-dimensional convolutional network, and step 103 includes:

[0078] Step 201: Perform multiple downsampling feature extraction on the voxel point cloud data based on multiple convolutional layers in the first feature network to obtain the three-dimensional positioning feature.

[0079] Herein, downsampling processing may refer to a processing method of screening and extracting target features according to specific feature recognition rules.

[0080] Exemplarily, the terminal may utilize multiple convolutional layers in the first feature network to perform multiple downsampling feature extractions on the voxel point cloud data, thereby obtaining the final three-dimensional positioning features.

[0081] Step 203: Through a preset voxel feature capture network, feature extraction is performed on the voxel point cloud data in each convolution layer to obtain the target recognition feature.

[0082] Exemplarily, the terminal may utilize a voxel feature capture network in each convolutional layer in the first feature network to perform feature extraction on voxel point cloud data that conforms to preset rules, thereby obtaining target recognition features.

[0083] In this embodiment, multiple layers of convolutional layers are used to perform multiple downsampling processes on the voxel point cloud data to obtain the final three-dimensional positioning features and target recognition features, which helps to improve the extraction effect of feature extraction on the point cloud data.

[0084] In one embodiment, Figure 3 As shown, step 107 includes:

[0085] Step 301: Acquire a central target area corresponding to the target to be measured according to the two-dimensional positioning feature.

[0086] The central target area may refer to the positioning center of the three-dimensional detection frame corresponding to the target to be detected, which may be a center point or a central area of ​​a preset size.

[0087] Exemplarily, the terminal may determine a central target area corresponding to the target to be measured in a preset two-dimensional plane according to the two-dimensional positioning feature.

[0088] Step 303: Determine a distance parameter between the central target area and the edge of the target to be measured, and obtain the positioning parameter according to the central target area and the distance parameter.

[0089] Exemplarily, the terminal may determine a distance parameter between a central target area and an edge of a target to be measured according to the two-dimensional positioning parameter, and obtain a positioning parameter according to the central target area and the distance parameter.

[0090] In this embodiment, the positioning parameters of the final three-dimensional detection frame are determined by the central area, which helps to improve the calculation efficiency of the positioning parameters and reduce the amount of calculation.

[0091] In one embodiment, Figure 4 As shown, step 303 includes:

[0092] Step 401: Generate several groups of candidate positioning parameters according to the central target area and the corresponding distance parameters.

[0093] For example, when the terminal calculates the positioning parameters, due to the information loss problem when projecting onto a two-dimensional plane, multiple sets of positioning parameters may be calculated. In this case, the obtained multiple sets of positioning parameters may be determined as candidate positioning parameters.

[0094] Step 403: Calculate the loss values ​​between the candidate positioning parameters and the preset reference true values ​​respectively, and determine the positioning parameters according to the loss values.

[0095] For example, the terminal may respectively calculate the loss value between the candidate positioning parameters and a preset reference ground truth, so that the terminal may determine the final positioning parameter to be adopted according to the obtained loss value.

[0096] In this embodiment, multiple groups of candidate positioning parameters are screened and determined by using loss values, which helps to improve the accuracy of selecting candidate positioning parameters.

[0097] In one embodiment, the positioning parameters may include the category, center position parameter, size parameter, and direction parameter of the three-dimensional detection box.

[0098] In one embodiment, Figure 5 As shown, step 107 includes:

[0099] Step 501: Acquire the positioning parameters through a preset detection head network, where the detection head network is an anchor-free network based on a center point.

[0100] Exemplarily, when obtaining positioning parameters through the detection head network, a center point-based anchor-free network (PVCP-RCNN model) can be used.

[0101] In this embodiment, the center-based anchor-free detection head network is used to help improve the efficiency of obtaining positioning parameters.

[0102] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0103] Based on the same inventive concept, the embodiment of the present application also provides a three-dimensional object detection device for implementing the three-dimensional object detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of a three-dimensional object detection device provided below can refer to the limitations of a three-dimensional object detection method above, and will not be repeated here.

[0104] In one embodiment, Figure 6 As shown, a three-dimensional target detection device is provided, including a voxel processing module, a feature extraction module, a feature projection module, a detection and positioning module and a target optimization module, wherein:

[0105] A voxel processing module is used to obtain point cloud data of the target to be detected in the detection space, and voxelize the point cloud data to obtain voxel point cloud data;

[0106] A feature extraction module, used to extract features from the voxel point cloud data using a pre-built first feature network, and obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data;

[0107] A feature projection module, used to project the target to be measured into a preset two-dimensional plane based on the three-dimensional positioning feature, so as to obtain the two-dimensional positioning feature of the target to be measured;

[0108] A detection and positioning module, used to obtain positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features;

[0109] The target optimization module is used to optimize the three-dimensional detection frame based on the target recognition feature to obtain a target detection result.

[0110] In one embodiment, the first feature network is a sparse three-dimensional convolutional network, and the feature extraction module includes:

[0111] A multi-level downsampling module, used for performing multiple downsampling feature extraction on the voxel point cloud data based on multiple convolutional layers in the first feature network to obtain the three-dimensional positioning feature;

[0112] The voxel feature module is used to extract features of the voxel point cloud data in each convolution layer through a preset voxel feature capture network to obtain the target recognition features.

[0113] In one embodiment, the detection and positioning module includes:

[0114] A center detection module, used for acquiring a center target area corresponding to the target to be detected according to the two-dimensional positioning feature;

[0115] The boundary distance module is used to determine the distance parameter between the central target area and the edge of the target to be measured, and obtain the positioning parameter according to the central target area and the distance parameter.

[0116] In one embodiment, the boundary distance module includes:

[0117] A candidate parameter module, used to generate several groups of candidate positioning parameters according to the central target area and the corresponding distance parameters;

[0118] The screening and determination module is used to respectively calculate the loss values ​​between the candidate positioning parameters and the preset reference true values, and determine the positioning parameters according to the loss values.

[0119] In one of the embodiments, the positioning parameters include the category, center position parameter, size parameter and direction parameter of the three-dimensional detection box.

[0120] In one embodiment, the detection and positioning module further includes:

[0121] The detection head module is used to obtain the positioning parameters through a preset detection head network, and the detection head network is an anchor-free frame network based on a center point.

[0122] Each module in the above-mentioned three-dimensional object detection device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module above.

[0123] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in FIG7. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected via a system bus, and the communication interface, the display unit, and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a three-dimensional target detection method is implemented. The display unit of the computer device is used to form a visually visible picture, which may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0124] Those skilled in the art will understand that the structure shown in FIG. 7 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0125] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0126] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0127] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0128] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0129] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0130] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A three-dimensional target detection method, characterized in that: The method comprises: Acquire point cloud data of the target to be detected in the detection space, and voxelize the point cloud data to obtain voxel point cloud data; Using a pre-constructed first feature network to perform feature extraction on the voxel point cloud data, and obtaining three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data; Based on the three-dimensional positioning feature, the target to be measured is projected into a preset two-dimensional plane to obtain the two-dimensional positioning feature of the target to be measured; Obtaining positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features; The three-dimensional detection frame is optimized based on the target recognition feature to obtain a target detection result.

2. The method according to claim 1, characterized in that The first feature network is a sparse three-dimensional convolutional network, and the use of the pre-constructed first feature network to extract features from the voxel point cloud data to obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data includes: Performing multiple downsampling feature extraction on the voxel point cloud data based on multiple convolutional layers in the first feature network to obtain the three-dimensional positioning feature; Through a preset voxel feature capture network, feature extraction is performed on the voxel point cloud data in each convolution layer to obtain the target recognition feature.

3. The method according to claim 1, characterized in that The obtaining of the positioning parameters of the three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features includes: Acquire a central target area corresponding to the target to be measured according to the two-dimensional positioning feature; A distance parameter between the central target area and the edge of the target to be measured is determined, and the positioning parameter is obtained according to the central target area and the distance parameter.

4. The method according to claim 3, characterized in that The determining of the distance parameter between the central target area and the edge of the target to be measured, and obtaining the positioning parameter according to the central target area and the distance parameter comprises: Generate several groups of candidate positioning parameters according to the central target area and the corresponding distance parameters; The loss values ​​between the candidate positioning parameters and the preset reference true values ​​are respectively calculated, and the positioning parameters are determined according to the loss values.

5. The method according to any one of claims 1 to 4, characterized in that: The positioning parameters include the category, center position parameter, size parameter and direction parameter of the three-dimensional detection frame.

6. The method according to claim 3, characterized in that The obtaining of the positioning parameters of the three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features also includes: The positioning parameters are obtained through a preset detection head network, and the detection head network is an anchor-free frame network based on a center point.

7. A three-dimensional target detection device, characterized in that: The device comprises: A voxel processing module is used to obtain point cloud data of the target to be detected in the detection space, and voxelize the point cloud data to obtain voxel point cloud data; A feature extraction module, used to extract features from the voxel point cloud data using a pre-built first feature network, and obtain three-dimensional positioning features and target recognition features corresponding to the voxel point cloud data; A feature projection module, used to project the target to be measured into a preset two-dimensional plane based on the three-dimensional positioning feature, so as to obtain the two-dimensional positioning feature of the target to be measured; A detection and positioning module, used to obtain positioning parameters of a three-dimensional detection frame corresponding to the target to be detected based on the two-dimensional positioning features; The target optimization module is used to optimize the three-dimensional detection frame based on the target recognition feature to obtain a target detection result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images