Laser radar target detection method and device based on deep learning
By projecting laser point cloud data onto a two-dimensional grid and applying deep learning techniques, the method addresses computational complexity issues, enabling real-time target detection in autonomous driving systems.
Patent Information
- Application Number
- CN202210051492.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-17
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2042-01-17
AI Technical Summary
Existing laser radar target detection methods in autonomous driving face challenges due to high computational complexity and inability to meet real-time processing requirements, as they involve large data processing volumes and complex structures, limiting their application in autonomous driving systems.
A method that projects laser point cloud data onto a two-dimensional grid plane, determining valid grids and their attributes, and using deep learning to extract grid features, similar to image deep learning techniques, thereby reducing computational load and enhancing real-time processing capabilities.
This approach maximizes network performance while minimizing point cloud information loss, effectively meeting real-time detection requirements in autonomous driving by leveraging image deep learning advancements.
Smart Images
Figure CN114419567B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of autonomous driving technology, and in particular to a deep learning-based laser radar target detection method and device in an intelligent driving system. Background Art
[0002] With the rapid development of intelligent driving technology in recent years, environmental perception is an important part of the intelligent driving system. Obstacle detection is a basic function in the environmental perception system. Obstacle types mainly include pedestrians, vehicles, bicycles and other common objects on structured roads. LiDAR can directly detect the real distance and size of the target and is the main sensor in the autonomous driving solution. As deep learning technology becomes more and more mature, deep learning technology is also playing an increasingly important role in the field of autonomous driving, and is gradually becoming the core technology in the field of autonomous driving.
[0003] Unlike cameras, LiDAR can directly measure the real size and distance of a target. However, the point cloud data generated by LiDAR is disordered and sparser than images. The density of point cloud data decreases with increasing distance. In view of the sparsity and disorder of LiDAR point cloud data, the processing methods of LiDAR are roughly divided into two categories: point coding and grid coding.
[0004] Point coding refers to directly processing the original point cloud data, automatically encoding the features of the point cloud through a deep learning network, and gradually expanding the network's field of view as the network deepens, extracting features around the points, and thus extracting various attributes of the point cloud. Point coding was first proposed by Charles R. Qi, PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation in 2017. Point coding is a technology that can not only be used as a coding layer for target detection tasks, but also complete other tasks such as classification and segmentation. However, in point coding technology, each point cloud needs to be processed, and the domain relationship of each point cloud needs to be calculated, so the amount of calculation is large, which is not suitable for the field of autonomous driving that requires real-time calculation.
[0005] Raster coding is involved in Yin Zhou, VoxelNet: End-to-End Learning for Point CloudBased 3D Object Detection, Martin Simon, Complex-YOLO: An Euler-Region-Proposal for Real-time 3D Object Detection on Point Clouds, Bin Yang, PIXOR: Real-time 3D Object Detection from Point Clouds, and Yan, SECOND: Sparsely Embeded Convolutional Detection. The raster coding method does not directly process point cloud data, but projects the three-dimensional point cloud into a spatial grid and then processes the points in the spatial grid. However, this processing method still has a large amount of processing power and cannot achieve real-time computing capabilities, which cannot meet the requirements of autonomous driving deployment.
[0006] Therefore, the existing methods have large data processing volume and complex structure, which cannot meet the requirements of real-time processing; and the structure is relatively complex and limited, and it is impossible to apply the network structure and research results related to the image deep learning neighborhood, and cannot meet the real-time requirements in the field of autonomous driving. Summary of the invention
[0007] The purpose of the embodiments of the present invention is to provide a laser radar target detection method based on deep learning to solve the problems existing in the prior art.
[0008] In order to solve the problems existing in the prior art, in a first aspect, the present invention provides a laser radar target detection method based on deep learning, the method comprising:
[0009] Acquire a laser point cloud; the laser point cloud includes a plurality of points, each point including three-dimensional coordinates in a first coordinate system;
[0010] After projecting the laser point cloud onto a plurality of sub-grids of a plane grid, determining valid grids in the sub-grids, and determining grid properties of the valid grids; the grid properties include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is a sub-grid with projection points; the average value point is a point corresponding to the average value of the projection points in each of the valid grids;
[0011] Input the three-dimensional coordinates of the point corresponding to each projection point in the effective grid in the first coordinate system, the distance between each projection point and the center point of the effective grid to which it belongs, and the distance between each projection point and the average value point of the effective grid to which it belongs into the deep learning network model to obtain the expanded grid attributes;
[0012] The expanded grid attributes are continuously input into the deep learning network to output a plurality of detection frames, each of which has a confidence level;
[0013] According to the confidence level, a target detection frame is determined.
[0014] In a possible implementation manner, after projecting the laser point cloud onto a plurality of sub-grids of a plane grid, determining a valid grid in the sub-grids specifically includes:
[0015] Taking the origin projection of the laser radar coordinate system as the center point of the plane grid, and determining the number of sub-grids in the plane grid according to a preset resolution of the plane grid;
[0016] It is determined whether the sub-grid has a projection point, and when the sub-grid has a projection point, the sub-grid is determined to be a valid grid.
[0017] In a possible implementation manner, determining the grid attribute of the valid grid specifically includes:
[0018] Determine the coordinates of the projection point in each valid grid according to the three-dimensional coordinates of the point corresponding to the projection point in the first coordinate system;
[0019] Determine the distance between each projection point and the center point of the valid grid to which it belongs according to the coordinates of the projection point;
[0020] Determine the average value point of each of the valid grids according to the coordinates of the projection points;
[0021] The distance between each projection point and the average value point of the valid grid to which it belongs is determined according to the coordinates of the projection point and the average value point.
[0022] In a possible implementation manner, determining the target detection frame according to the confidence level specifically includes:
[0023] Determine the target detection box based on the confidence and non-maximum suppression NMS algorithm.
[0024] In a second aspect, the present invention provides a laser radar target detection device based on deep learning, the device comprising:
[0025] An acquisition module, the acquisition module is used to acquire a laser point cloud; the laser point cloud includes a plurality of points, each point includes a three-dimensional coordinate in a first coordinate system;
[0026] A grid determination module, the determination module is used to determine the valid grids in the sub-grids after projecting the laser point cloud onto multiple sub-grids of the plane grid, and determine the grid attributes of the valid grids; the grid attributes include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is a sub-grid with projection points; the average value point is a point corresponding to the average value of the projection points in each of the valid grids;
[0027] A first learning module, the learning module is used to input the three-dimensional coordinates of the point corresponding to each projection point in the effective grid in the first coordinate system, the distance of each projection point from the center point of the effective grid to which it belongs, and the distance of each projection point from the average value point of the effective grid to which it belongs into the deep learning network model to obtain the expanded grid attributes;
[0028] A second learning module, wherein the second learning module is used to continue to input the expanded grid attributes into a deep learning network and output a plurality of detection boxes; each detection box has a confidence level;
[0029] The target detection frame determination module is used to determine the target detection frame according to the confidence level.
[0030] A third aspect of the present invention provides a computer server, comprising: a memory, a processor and a transceiver;
[0031] The processor is used to couple with the memory, read and execute instructions in the memory, so as to implement the laser radar target detection method as described in the first aspect;
[0032] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.
[0033] The fourth aspect of the present invention provides a chip system, including a processor, the processor is coupled to a memory, the memory stores program instructions, and when the program instructions stored in the memory are executed by the processor, the laser radar target detection method as described in the first aspect is implemented.
[0034] A fifth aspect of the present invention provides a computer system, comprising a memory, and one or more processors communicatively connected to the memory;
[0035] The memory stores instructions that can be executed by the one or more processors, and the instructions are executed by the one or more processors so that the one or more processors implement the laser radar target detection method as described in the first aspect.
[0036] A sixth aspect of the present invention provides a mobile tool, comprising the computer server described in the third aspect above.
[0037] The deep learning-based laser radar target detection method provided in this application projects the point cloud data of the three-dimensional space into the two-dimensional grid plane, and calculates the grid features of the point cloud in units of effective grids. This method is more flexible to use, and the processing method is similar to the image deep learning method. After years of development, the field of image deep learning has accumulated a lot of experience. By using the relevant achievements in the field of image deep learning, the upper limit of network performance can be improved, and the real-time processing requirements of autonomous driving can be met. The traditional convolutional network is used to extract grid features to achieve the purpose of detecting the target. This achieves the goal of maximizing network performance while reducing the loss of point cloud information, and meeting the requirements of real-time detection of autonomous driving. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 A schematic diagram of a process flow of a laser radar target detection method based on deep learning provided in Embodiment 1 of the present invention;
[0039] Figure 2 for Figure 1 Specific flow chart of step 120;
[0040] Figure 3 Schematic diagram of LiDAR and plane grid;
[0041] Figure 4 A schematic diagram of the structure of a deep learning-based laser radar target detection device provided in Example 2 of the present invention. DETAILED DESCRIPTION
[0042] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.
[0043] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0044] Embodiment 1
[0045] Figure 1 A schematic diagram of a deep learning-based laser radar target detection method according to the first embodiment of the present invention is shown in FIG. Figure 1 As shown, the method comprises the following steps:
[0046] Step 110, obtaining a laser point cloud; the laser point cloud includes a plurality of points, each point including three-dimensional coordinates in a first coordinate system.
[0047] Specifically, in the field of autonomous driving, for example, a laser radar is installed on an autonomous driving vehicle, and the laser radar can obtain a laser point cloud in real time. The laser point cloud is a series of disordered spatial coordinate points. The three-dimensional coordinate refers to the position and height of each point in the laser point cloud, which can be expressed as (x, y, h), where x represents the horizontal axis coordinate, y represents the vertical axis coordinate, and h represents the height. The first coordinate system can be the laser radar coordinate system.
[0048] Step 120, after projecting the laser point cloud onto a plurality of sub-grids of the plane grid, determine the valid grids in the sub-grids, and determine the grid attributes of the valid grids; the grid attributes include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is a sub-grid with projection points; the average value point is a point corresponding to the average value of the projection points in each valid grid.
[0049] The second coordinate system is a plane coordinate system, and the plane here may be a ground plane. The coordinates of a point in the first coordinate system may be converted into coordinates in the second coordinate system through coordinate system conversion.
[0050] Figure 2 for Figure 1 After projecting the laser point cloud onto a plurality of sub-grids of the plane grid in step 120, determining a valid grid in the sub-grid specifically includes the following steps:
[0051] Step 1201, taking the origin of the laser radar coordinate system as the center point of the plane grid, and determining the number of sub-grids in the plane grid according to a preset resolution of the plane grid;
[0052] The plane grid can be a 2-D grid. When projecting the laser point cloud onto the plane grid, the center point of the laser point cloud can be projected onto the origin of the grid. For example, taking the origin of the laser radar coordinate system as the origin, taking the points within 60 meters in front, behind, left and right of the laser radar coordinate system, setting the grid resolution to 0.25m (i.e., the side length of the sub-grid is 0.25m), the number of sub-grids is 480*480 (see Appendix). Figure 3 The number of sub-grids shown is for example only).
[0053] Step 1202, determining whether there is a projection point in the sub-grid, and when there is a projection point in the sub-grid, determining that the sub-grid is a valid grid.
[0054] Among them, when there are projection points in the sub-grids, these sub-grids can be determined as valid grids, and the grids without projection points can be determined as invalid grids. In subsequent processing, only valid grids can be processed, and invalid grids are not processed, thereby greatly reducing the subsequent calculation amount.
[0055] Figure 2 Steps 1203-1206 in are specific steps for determining the grid properties of the valid grid in step 120, as shown below:
[0056] Step 1203, determining the coordinates of the projection point according to the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system;
[0057] Specifically, there is a conversion relationship between the first coordinate system and the second coordinate system, based on which the coordinates of each projection point can be calculated, and the coordinates of the projection point refer to the x-coordinate and the y-coordinate of the projection point.
[0058] Step 1204, determining the distance between each projection point and the center point of the valid grid to which it belongs according to the coordinates of the projection point;
[0059] Specifically, since each projection point has coordinates, the distance between each projection point and the center point of the finite grid to which it belongs can be determined.
[0060] Step 1205, determining the average value point of each valid grid according to the coordinates of the projection point;
[0061] The average value point refers to the sum of the x-coordinate values of all projection points in each valid grid divided by the number of projection points, and the sum of the y-coordinate values divided by the number of projection points.
[0062] Step 1206, determining the distance between each projection point and the average value point of the valid grid to which it belongs based on the coordinates of the projection point and the average value point.
[0063] Specifically, since each projection point has coordinates, the distance between each projection point and the average value point of the valid grid to which it belongs is calculated.
[0064] Subsequently, the x-coordinate, y-coordinate, and h-coordinate of each point, the x-distance of the projection point from the center point of the grid, the y-distance of the projection point from the center point of the grid, the x-distance of the projection point from the average point in each valid grid, and the y-distance of the projection point from the average point in each valid grid can be used as grid attributes, thereby realizing point coding of the projection points in the grid. Thus, even if the number of projection points included in the grid is not fixed, each point has a 7-channel grid feature after point coding.
[0065] Step 130, the three-dimensional coordinates of the point corresponding to each projection point in the valid grid in the first coordinate system, the distance of each projection point from the center point of the valid grid to which it belongs, and the distance of each projection point from the average value point of the valid grid to which it belongs are input into the deep learning network model to obtain the expanded grid attributes.
[0066] Specifically, the raster attributes are input into the learning network for learning. In one example, 7 values in the raster attributes are input into the deep learning network model, and 64-channel raster features are output. The deep learning network model here is a commonly used deep learning network model and will not be described in detail here.
[0067] By way of example and not limitation, the raster output channels are not limited to 64 layers.
[0068] Step 140, the expanded raster attributes are further input into the deep learning network, and a plurality of detection boxes are output; each detection box has a confidence level.
[0069] Continuing with the previous example, the expanded raster attributes are input into the Region Proposal Networks (RPN) layer of the feature pyramid networks (FPN), which outputs multiple detection boxes; each detection box has a confidence level.
[0070] Specifically, after the 64-channel grid features are input into the RPN layer of FPN, multiple detection boxes can be output, each of which has a position and confidence, and the confidence can take a value in (0,1).
[0071] Among them, the RPN output is not limited to 2 outputs, and different branches and network structures can be designed according to actual needs.
[0072] Step 150: determine the target detection frame according to the confidence level.
[0073] Specifically, the number of detection frames output by RPN is large, but only a few of them are valid, so they need to be processed to merge adjacent detection frames and filter invalid detection frames. Here, the detection frames can be processed by nonmaximum suppression (NMS) to determine the target detection frame.
[0074] The deep learning-based laser radar target detection method provided in this application projects the point cloud data of the three-dimensional space into the two-dimensional grid plane, and calculates the grid features of the point cloud in units of effective grids. This method is more flexible to use, and the processing method is similar to the image deep learning method. After years of development, the field of image deep learning has accumulated a lot of experience. By using the relevant achievements in the field of image deep learning, the upper limit of network performance can be improved, and the real-time processing requirements of autonomous driving can be met. The traditional convolutional network is used to extract grid features to achieve the purpose of detecting the target. This achieves the goal of maximizing network performance while reducing the loss of point cloud information, and meeting the requirements of real-time detection of autonomous driving.
[0075] Embodiment 2
[0076] Figure 4 This is a schematic diagram of the structure of a laser radar target detection device based on deep learning provided in the second embodiment of the present invention. Figure 3 As shown, the deep learning-based laser radar target detection device includes: an acquisition module 410, a grid determination module 420, a first learning module 430, a second learning module 440 and a target detection frame determination module 450.
[0077] The acquisition module 410 is used to acquire a laser point cloud; the laser point cloud includes a plurality of points, each point includes a three-dimensional coordinate in a first coordinate system;
[0078] The grid determination module 420 is used to determine the valid grids in the sub-grids after projecting the laser point cloud onto multiple sub-grids of the plane grid, and determine the grid attributes of the valid grids; the grid attributes include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is the sub-grid with the projection point; the average value point is the point corresponding to the average value of the projection points in each valid grid;
[0079] The first learning module 430 is used to input the three-dimensional coordinates of the point corresponding to each projection point in the valid grid in the first coordinate system, the distance between each projection point and the center point of the valid grid to which it belongs, and the distance between each projection point and the average value point of the valid grid to which it belongs into the deep learning network model to obtain the expanded grid attributes;
[0080] The second learning module 440 is used to continue to input the expanded grid attributes into the deep learning network and output multiple detection boxes; each detection box has a confidence level;
[0081] The target detection frame 450 is used for determining the module to determine the target detection frame according to the confidence level.
[0082] The acquisition module 410 implements the same function as step 110 in the first embodiment. For specific implementation, please refer to the relevant description of step 110 in the first embodiment, which will not be repeated here.
[0083] The first learning module 430 implements the same function as step 130 in embodiment 1. For specific implementation, please refer to the relevant description of step 130 in embodiment 1, which will not be repeated here.
[0084] The second learning module 440 implements the same function as step 140 in embodiment 1. For specific implementation, please refer to the relevant description of step 140 in embodiment 1, which will not be repeated here.
[0085] After the grid determination module 420 projects the laser point cloud onto multiple sub-grids of the plane grid, determining the valid grids in the sub-grids specifically includes: projecting the origin of the laser radar coordinate system as the center point of the plane grid, and determining the number of sub-grids in the plane grid according to a preset resolution of the plane grid; judging whether there are projection points in the sub-grid, and when there are projection points in the sub-grid, determining the sub-grid to be a valid grid.
[0086] The grid determination module 420 determines the grid properties of the valid grid specifically by: determining the coordinates of the projection point according to the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system; determining the distance of each projection point from the center point of the valid grid to which it belongs according to the coordinates of the projection point; determining the average value point of each valid grid according to the coordinates of the projection point; determining the distance of each projection point from the average value point of the valid grid to which it belongs according to the coordinates of the projection point and the average value point.
[0087] The target detection frame determination module 450 determines the target detection frame according to the confidence level, specifically including: determining the target detection frame according to the confidence level and the non-maximum suppression (NMS) algorithm.
[0088] Embodiment 3
[0089] Embodiment 3 of the present invention provides a chip system, including a processor, a coupling between the processor and a memory, the memory storing program instructions, and when the program instructions stored in the memory are executed by the processor, any one of the laser radar target detection methods provided in embodiment 1 is implemented.
[0090] Embodiment 4
[0091] Embodiment 4 of the present invention provides a computer server, including: a memory, a processor and a transceiver;
[0092] The processor is used to couple with the memory, read and execute instructions in the memory, so as to implement any one of the laser radar target detection methods provided in the first embodiment above;
[0093] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.
[0094] Embodiment 5
[0095] Embodiment 5 of the present invention provides a computer-readable storage medium, including a program or an instruction. When the program or the instruction is executed on a computer, any one of the laser radar target detection methods provided in Embodiment 1 is implemented.
[0096] Embodiment 6
[0097] Embodiment 6 provides a computer program product comprising instructions. When the computer program product is run on a computer, the computer is enabled to execute any one of the laser radar target detection methods provided in Embodiment 1.
[0098] Embodiment 7
[0099] Embodiment 7 of the present invention provides a mobile tool, including the above-mentioned computer server.
[0100] The mobile tool can be any movable tool, such as vehicles (such as vacuum trucks, sweepers, floor scrubbers, logistics carts, passenger cars, sanitation vehicles, buses, coaches, vans, trucks, trucks, trailers, trailers, cranes, excavators, bulldozers, road trains, sweepers, sprinkler trucks, garbage trucks, engineering vehicles, rescue vehicles, AGV (Automated Guided Vehicle), etc.), motorcycles, bicycles, tricycles, carts, robots, sweepers, balance bikes, etc. The present application does not strictly limit the types of mobile tools and they are not listed here in detail.
[0101] The professionals should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0102] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0103] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A laser radar target detection method based on deep learning, characterized in that: The method comprises: Acquire a laser point cloud; the laser point cloud includes a plurality of points, each point including three-dimensional coordinates in a first coordinate system; After projecting the laser point cloud onto a plurality of sub-grids of a plane grid, determining valid grids in the sub-grids, and determining grid properties of the valid grids; the grid properties include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is a sub-grid with projection points; the average value point is a point corresponding to the average value of the projection points in each of the valid grids; Input the three-dimensional coordinates of the point corresponding to each projection point in the effective grid in the first coordinate system, the distance between each projection point and the center point of the effective grid to which it belongs, and the distance between each projection point and the average value point of the effective grid to which it belongs into the deep learning network model to obtain the expanded grid attributes; The expanded grid attributes are continuously input into the deep learning network to output a plurality of detection frames, each of which has a confidence level; According to the confidence level, a target detection frame is determined.
2. The method according to claim 1, characterized in that After projecting the laser point cloud onto a plurality of sub-grids of a plane grid, determining a valid grid in the sub-grids specifically includes: Taking the origin projection of the laser radar coordinate system as the center point of the plane grid, and determining the number of sub-grids in the plane grid according to a preset resolution of the plane grid; It is determined whether the sub-grid has a projection point, and when the sub-grid has a projection point, the sub-grid is determined to be a valid grid.
3. The method according to claim 1, characterized in that: The step of determining the grid attributes of the valid grid specifically includes: Determine the coordinates of the projection point in each valid grid according to the three-dimensional coordinates of the point corresponding to the projection point in the first coordinate system; Determine the distance between each projection point and the center point of the valid grid to which it belongs according to the coordinates of the projection point; Determine the average value point of each of the valid grids according to the coordinates of the projection points; The distance between each projection point and the average value point of the valid grid to which it belongs is determined according to the coordinates of the projection point and the average value point.
4. The method according to claim 1, characterized in that: Determining the target detection frame according to the confidence level specifically includes: Determine the target detection box based on the confidence and non-maximum suppression NMS algorithm.
5. A laser radar target detection device based on deep learning, characterized in that: The device comprises: An acquisition module, the acquisition module is used to acquire a laser point cloud; the laser point cloud includes a plurality of points, each point includes a three-dimensional coordinate in a first coordinate system; A grid determination module, the grid determination module is used to determine the valid grids in the sub-grids after projecting the laser point cloud onto multiple sub-grids of the plane grid, and determine the grid attributes of the valid grids; the grid attributes include the three-dimensional coordinates of the point corresponding to the projection point in each valid grid in the first coordinate system, the distance from each projection point to the center point of the valid grid to which it belongs, and the distance from each projection point to the average value point of the valid grid to which it belongs; the valid grid is a sub-grid with projection points; the average value point is a point corresponding to the average value of the projection points in each of the valid grids; A first learning module, wherein the first learning module is used to input the three-dimensional coordinates of the point corresponding to each projection point in the valid grid in the first coordinate system, the distance of each projection point from the center point of the valid grid to which it belongs, and the distance of each projection point from the average value point of the valid grid to which it belongs into the deep learning network model to obtain the expanded grid attributes; A second learning module, wherein the second learning module is used to continue to input the expanded grid attributes into a deep learning network and output a plurality of detection boxes; each detection box has a confidence level; The target detection frame determination module is used to determine the target detection frame according to the confidence level.
6. A computer server, characterized in that: include: memory, processors, and transceivers; The processor is used to couple with the memory, read and execute instructions in the memory, so as to implement the deep learning-based laser radar target detection method according to any one of claims 1 to 4; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.
7. A chip system, characterized in that: It includes a processor, the processor is coupled to a memory, the memory stores program instructions, and when the program instructions stored in the memory are executed by the processor, the deep learning-based laser radar target detection method described in any one of claims 1 to 4 is implemented.
8. A computer system, characterized in that: comprising a memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors, and the instructions are executed by the one or more processors so that the one or more processors implement the deep learning-based lidar target detection method as described in any one of claims 1-4.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is executed by the processor to implement the deep learning-based laser radar target detection method described in any one of claims 1 to 4.
10. A mobile tool, characterized in that: The computer server comprising the above-mentioned claim 6.