Patrol detection robot mapping navigation motion control method based on laser radar

Abstract

The invention relates to the field of mobile inspection robots, in particular to an inspection detection robot mapping navigation motion control method based on a laser radar. Comprising the steps that IMU data in the motion process of the patrol detection robot and 3D point cloud data obtained through 3D laser radar scanning are acquired in real time; performing closed-loop fusion on the 3D point cloud data and IMU data and a backward calculation process, and compensating radar sampling mapping to obtain a 3D point cloud map; detecting obstacles encountered in the movement process in real time based on the 3D point cloud data, and updating the obstacles to a 2D map converted based on the 3D point cloud map; and performing motion obstacle avoidance control on the patrol detection robot based on the updated 2D map. According to the method, distortion caused by robot movement can be eliminated, accurate mapping can be achieved, and therefore patrol detection robot mapping navigation and movement obstacle avoidance control can be well achieved.

Description

Technical Field

[0001] The present invention relates to the field of mobile inspection robots, and in particular to a laser radar-based patrol inspection robot mapping navigation motion control method. Background Art

[0002] With the advancement of science and technology, inspection robots are widely used in industrial production, security, military and other fields. In industrial production, inspection robots can perform automated inspections to ensure equipment safety and normal operation of production lines, and take prompt measures in the event of emergencies. At the same time, inspection robots also play an important role in security and military fields, gradually replacing manual work to complete security and inspection tasks in dangerous scenarios in industrial parks, providing safety guarantees for normal production operations in factories.

[0003] High-precision security inspection robots rely on sensors such as vision or lidar for environmental perception to achieve positioning and navigation functions. Compared with visual sensors, lidar is not affected by ambient light and can maintain high-precision positioning and recognition in extreme weather conditions such as night, rain or strong light.

[0004] However, existing LiDAR inspection robots have poor recognition capabilities for scene changes during autonomous navigation. When encountering dynamic scene changes, such as sudden intrusion of pedestrians or vehicles, the robot cannot quickly perceive and respond, often leading to increased collision risks. At the same time, when complex scenes change, it is difficult for the robot to quickly recognize the new spatial structure and it still follows the old path, resulting in being trapped or unable to complete the inspection task.

[0005] Therefore, there is an urgent need for a laser radar-based patrol inspection robot mapping navigation motion control method, which enables the patrol inspection robot to automatically map and navigate and achieve motion obstacle avoidance control. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a laser radar-based patrol detection robot mapping navigation motion control method, which can eliminate the distortion caused by the robot's movement and accurately build maps, thereby well realizing the patrol detection robot's mapping navigation and motion obstacle avoidance control.

[0007] In order to solve the above technical problems, the technical solution of the present invention is: a laser radar-based patrol detection robot mapping navigation motion control method, comprising:

[0008] Real-time acquisition of IMU data during the movement of the patrol inspection robot and 3D point cloud data obtained through 3D laser radar scanning;

[0009] The 3D point cloud data is closed-loop fused with IMU data and the backward calculation process, and the radar sampling is compensated to build the map to obtain a 3D point cloud map;

[0010] Detect obstacles encountered during movement in real time based on 3D point cloud data, and update the obstacles to the 2D map converted from the 3D point cloud map;

[0011] The patrol detection robot is controlled to avoid obstacles based on the updated 2D map.

[0012] Furthermore, the 3D point cloud data is closed-loop fused with the IMU data of the patrol detection robot and the backward calculation process, and the radar sampling and mapping are compensated to obtain a 3D point cloud map; specifically, it includes:

[0013] Extract features of edge points and plane points in 3D point cloud data;

[0014] Back-propagation motion compensation of 3D point cloud data based on the extracted features and the output of the forward propagation operation on the IMU data;

[0015] The output of the backward motion compensation and the output of the forward operation are subjected to residual calculation and state update, and then a judgment is made. If it is judged to be a new observation point, it is updated to the 3D point cloud map.

[0016] Further, a backward-forward motion compensation is performed based on the extracted features and the output of the forward-forward operation on the IMU data; specifically:

[0017] In 3D LiDAR k At the sampling time, for multiple feature points between two adjacent frames of IMU data, the left IMU frame is used as the reference, and then the backward calculation is performed to project the relative pose of the point obtained by the backward calculation to time t k time, and then perform coordinate transformation.

[0018] Furthermore, obstacles encountered during the movement are detected in real time based on the 3D point cloud data, and the obstacles are updated to the 3D point cloud map, including:

[0019] Dynamic sparse coding branches are used to dynamically encode 3D point cloud data, and multi-scale sampling is used to learn multi-scale features of point clouds.

[0020] Auxiliary supervision branches are used to alleviate the problem of spatial information blur and loss caused by downsampling of dynamic sparse coding branches;

[0021] Use the key point branch to sample the key points of multi-scale features to obtain key point features;

[0022] The ROI-grid pooling module is used to refine the proposals of multi-scale features and key point features, output the prediction results, and update the predicted obstacles to the 3D point cloud map.

[0023] Furthermore, the dynamic sparse coding branch includes four sparse convolution blocks connected in sequence, which sample the 3D point cloud data by 1, 2, 4, and 8 times in sequence to obtain multi-scale features;

[0024] The auxiliary supervision branch includes three upsampling blocks connected in sequence, corresponding to the last three sparse convolution blocks. The upsampling block first converts the voxel coordinates of the corresponding sparse convolution block into point coordinates in the real scene, then interpolates the sampling points through the feature propagation layer, and finally connects all point features to learn the structural knowledge characteristics of the target object;

[0025] The key point branch includes four voxel set abstract blocks connected in sequence. The voxels of the four sparse convolution blocks are sampled by the farthest sampling method to select multiple key points. Each key point encodes the features of its adjacent non-empty voxels, abstracts the global and local features of the set and summarizes them into a key point feature.

[0026] The present invention also provides a laser radar-based patrol detection robot mapping and navigation device, comprising:

[0027] The acquisition module is used to obtain the IMU data of the patrol detection robot during its movement and the 3D point cloud data obtained by 3D laser radar scanning in real time;

[0028] Mapping module, which is used to close the loop of 3D point cloud data, fuse IMU data and backward calculation process, compensate radar sampling and map, and obtain 3D point cloud map;

[0029] The obstacle detection and annotation module is used to detect obstacles encountered during movement in real time based on 3D point cloud data, and update the obstacles to the 2D map converted from the 3D point cloud map;

[0030] The motion obstacle avoidance control module is used to perform motion obstacle avoidance control on the patrol detection robot based on the updated 2D map.

[0031] The present invention also provides a patrol detection robot, which is configured to use a patrol detection robot mapping navigation motion control method based on laser radar for mapping navigation.

[0032] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the laser radar-based patrol detection robot mapping navigation motion control method are implemented.

[0033] The present invention also provides a computer program product, including a computer program, which implements the steps of a laser radar-based patrol detection robot mapping navigation motion control method when the computer program is executed by a processor.

[0034] After adopting the above technical scheme, the present invention integrates IMU data and backward calculation process to compensate radar sampling mapping, so as to realize the precise mapping and navigation tasks of the inspection robot; and based on the point cloud detection method of 3D laser radar, it performs real-time monitoring of obstacles encountered by the inspection robot during patrol, which is used for subsequent planning and design; and it also updates the obstacles detected by the 3D laser radar to the 2D map of the navigation, so as to realize real-time update of the map and thus realize obstacle avoidance motion control. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the laser radar-based patrol inspection robot mapping navigation motion control method of the present invention;

[0036] Figure 2 This is a framework diagram of the laser radar fusion IMU mapping (SLAM) of the present invention;

[0037] Figure 3 It is a framework diagram of the obstacle detection (KASNet network) of the present invention;

[0038] Figure 4 A framework diagram of the key point assisted supervision region proposal (KAS-RPN network) of the present invention;

[0039] Figure 5 A hardware system connection diagram of the patrol detection robot of the present invention;

[0040] Figure 6 It is the global 3D point cloud map of the present invention;

[0041] Figure 7 The 3D point cloud update map of the present invention;

[0042] Figure 8 This is a comparison diagram of the 2D point cloud before and after updating of the present invention; wherein,

[0043] Figure 7 In the figure, (a) is the 3D global image, and (b) is the 3D local image;

[0044] Figure 8 In the figure, (a) is the 2D image before updating, and (b) is the 2D image after updating. DETAILED DESCRIPTION

[0045] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0046] like Figures 1 to 4 As shown, a laser radar-based patrol detection robot mapping navigation motion control method includes:

[0047] Step S1, real-time acquisition of IMU data during the movement of the patrol detection robot and 3D point cloud data obtained by 3D laser radar scanning;

[0048] Step S2, close-loop fusion of 3D point cloud data with IMU data and backward calculation process, compensate radar sampling and mapping, and obtain a 3D point cloud map;

[0049] Step S3, detecting obstacles encountered during the movement in real time based on the 3D point cloud data, and updating the obstacles to the 2D map converted from the 3D point cloud map;

[0050] Step S4, performing motion obstacle avoidance control on the patrol detection robot based on the updated 2D map.

[0051] In this embodiment, if Figure 2 As shown, in step S2, the 3D point cloud data is closed-loop fused with the IMU data of the patrol detection robot and the backward calculation process, and the radar sampling is compensated to build a map to obtain a 3D point cloud map; specifically, it includes:

[0052] Step S21, extracting features of edge points and plane points in 3D point cloud data;

[0053] Specifically, this step is to first sample the laser radar point cloud data at intervals, and then send it to the point cloud feature extraction block. After downsampling the point cloud, the point cloud density is reduced and outliers are removed. After feature point extraction, the features of edge points and plane points are extracted by the curvature method. Specifically, the curvature method calculates the curvature of the curve formed by five points on the left and right of the same point, a total of eleven points. If the curvature is relatively large, the point is an edge point. If the curvature is relatively small, the point is a plane point. It can be judged by setting a threshold, and the filtered feature points are used for mapping and positioning.

[0054] Step S22, performing backward-forward motion compensation on the 3D point cloud data based on the extracted features and the output of the forward-forward operation on the IMU data;

[0055] Since the new scan time is at t k moment, but the feature points are at their respective sampling times ρ j The measurement is performed at the same time, which leads to motion distortion. k-1 The IMU measurement is used as the input for back-calculation to obtain the time ρ j A relative pose is generated between the end time of the laser radar scan, and then the pose is projected to the scan end measurement value to obtain the compensated point cloud. The crude state quantity and error covariance matrix are obtained through forward calculation, and the newly scanned feature points are fused with the results of forward calculation to obtain the optimal state update.

[0056] In other words, considering that the laser radar has a relatively poor effect on self-generated attitude measurement, this step is to use IMU data for real-time motion estimation and attitude estimation. By filtering the acceleration and angular velocity data of the IMU, the attitude and motion state of the robot can be estimated. By using the estimated attitude and motion state to perform motion compensation on the laser radar data, the distortion caused by the robot motion can be eliminated. That is, by fusing IMU data, the accuracy and robustness of attitude measurement can be improved and the problem of geometric environment structure degradation can be alleviated.

[0057] Step S23, the output of the backward motion compensation and the output of the forward calculation operation are iterated through iterative Kalman, that is, residual calculation and state update are performed first, and then judgment is made. Iterative Kalman filtering can reduce linear errors. If it is judged to be a new observation point, it is updated to the 3D point cloud map.

[0058] In this step, the state estimate is adjusted based on the calculated residual and Kalman gain. It can be roughly understood as integrating the current observed data into the previous state estimate.

[0059] Wherein, step S22 is specifically as follows:

[0060] In 3D LiDAR k At the sampling time, for multiple feature points between two adjacent frames of IMU data, the left IMU frame is used as the reference, and then the backward calculation is performed to project the relative pose of the point obtained by the backward calculation to time t k time, and then perform coordinate transformation.

[0061] That is, in step S2, in order to solve the problem of environmental geometry degradation, a tightly coupled iterative Kalman filter is used to fuse IMU measurements. The Kalman filter fusion algorithm is a commonly used data fusion algorithm that can fuse multiple sensor data to obtain a more accurate estimation result. Therefore, a tightly coupled iterative Kalman filter can be used to fuse IMU measurements to solve the problem of environmental geometry degradation.

[0062] The movement of the robot will cause distortion of the point cloud data. For example, rotation and translation will change the position of the point cloud in the ground coordinate system. The LiDAR data is compensated by fusing the IMU data. This is mainly done through back-calculation and motion compensation.

[0063] In this embodiment, the sparsity and occlusion of point clouds are challenging issues in 3D target detection. LiDAR point cloud data is sparsely distributed due to factors such as distance, occlusion, and posture, making it difficult to accurately capture the shape of the target. At the same time, occlusion between targets limits feature extraction and increases the difficulty of LiDAR detection.

[0064] Based on this, this embodiment proposes a new KASNet network to solve the problem of detecting long-distance density-unbalanced target point clouds in 3D detection. Step S3 is performed based on the KASNet network. Figure 3 As shown, step S3 specifically includes:

[0065] Step S31, using a dynamic sparse coding branch to perform dynamic voxel coding on the 3D point cloud data, and multi-scale sampling to learn multi-scale features of the point cloud;

[0066] Step S32, using the auxiliary supervision branch to alleviate the problem of spatial information blur and loss caused by downsampling of the dynamic sparse coding branch;

[0067] Step S33, using the key point branch to perform key point sampling on the multi-scale features to obtain key point features;

[0068] In step S34, the ROI-grid pooling module is used to refine the proposals of the multi-scale features and key point features, output the prediction results, and update the predicted obstacles to the 3D point cloud map.

[0069] Among them, Figure 4 As shown in the figure, the dynamic sparse coding branch, auxiliary supervision branch and key point branch together form the key point auxiliary supervision region proposal network (KAS-RPN), which performs dynamic encoding and generates high recall proposals to improve target positioning accuracy.

[0070] The dynamic sparse coding branch consists of four sparse convolution blocks (block1 to block4) connected in sequence, which sample the 3D point cloud data by 1, 2, 4, and 8 times in turn to obtain multi-scale features. These blocks process voxel features layer by layer and produce discriminative features with smaller resolution. The RPN head will be used for BEV planarization and generate candidate bounding boxes with high recall.

[0071] The auxiliary supervision branch is used to guide the model to learn spatial information, including three upsampling blocks connected in sequence, corresponding to the last three sparse convolution blocks respectively. The upsampling block first converts the voxel coordinates of the corresponding sparse convolution block into point coordinates in the real scene, that is, converts the features to points, and then interpolates the sampling points through the feature propagation layer. Finally, all point features are connected to learn the structural knowledge characteristics of the target object; and the auxiliary supervision branch task is abandoned in the reasoning stage, without increasing the additional calculation amount, which can effectively reduce the complexity of the model.

[0072] The key point branch is used for dynamic coding weighting, including four voxel set abstract blocks connected in sequence. The key point sampling method, namely the farthest sampling (FPS) algorithm, selects 2048 key points for the voxels of the four sparse convolution blocks. Each key point encodes the features of its adjacent non-empty voxels, abstracts the global and local features of the set and summarizes them into a key point feature. These key point features not only maintain accurate positions, but also encode rich scene contexts to improve 3D detection performance.

[0073] It should be noted that the KASNet network is a two-stage detection method. First, a first-stage prediction is performed through KAS-RPN. Secondly, the 3D detection box and key point features predicted in the first stage are sent to the ROI pooling stage to refine the detection box and output the final detection result.

[0074] In one embodiment, a laser radar-based patrol inspection robot mapping and navigation device includes:

[0075] The acquisition module is used to obtain the IMU data of the patrol detection robot during its movement and the 3D point cloud data obtained by 3D laser radar scanning in real time;

[0076] Mapping module, which is used to close the loop of 3D point cloud data, fuse IMU data and backward calculation process, compensate radar sampling and map, and obtain 3D point cloud map;

[0077] The obstacle detection and annotation module is used to detect obstacles encountered during movement in real time based on 3D point cloud data, and update the obstacles to the 2D map converted from the 3D point cloud map;

[0078] The motion obstacle avoidance control module is used to perform motion obstacle avoidance control on the patrol detection robot based on the updated 2D map.

[0079] In one embodiment, a patrol inspection robot is configured as the mapping and navigation of the patrol inspection robot mapping and navigation motion control method based on laser radar in the above embodiment.

[0080] In one embodiment, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the laser radar-based patrol inspection robot mapping navigation motion control method in the above embodiment.

[0081] In one embodiment, a computer program product includes a computer program, which, when executed by a processor, implements the steps of the laser radar-based patrol inspection robot mapping navigation motion control method in the above embodiment.

[0082] The hardware platform involved in the above embodiment is as follows Figure 5As shown in the figure, the connection between the laser radar and the hardware car, as well as the internal composition of the car, is demonstrated. The power supply supplies power to the motor controller, motor and Jetson Xavier respectively. Jetson Xavier controls the motor controller, and the motor controller controls the motor to drive the wheels to control the movement of the car.

[0083] The following is a detailed introduction to the laser radar-based patrol inspection robot mapping navigation motion control method involved in the above embodiment in combination with specific experimental analysis.

[0084] A: Experiment of compensating radar sampling and mapping by integrating IMU data and back-calculation process;

[0085] 1) Data preprocessing

[0086] After acquiring the lidar data, the feature points of the edges and surfaces are extracted through the SLAM technology, and the most useful feature points are selected for mapping and positioning.

[0087] 2) State Estimation

[0088] Since the sampling time of each point of the laser radar is different, and the laser radar is moving, it will cause motion distortion. Therefore, the data we want to obtain is the sampling of all points at the same time, that is, at t k Therefore, according to the IMU integral estimated pose, each point is transferred to t k Usually the frequency of the point is greater than the frequency of the IMU. Therefore, for multiple feature points between two adjacent IMU frames, the left IMU frame is used as the standard, and then the backward calculation is performed to project the relative pose of the point obtained by the backward calculation to time t k time, and then perform coordinate transformation.

[0089] 3) Map updates

[0090] After the Kalman filter iteration converges, the updated feature points are projected from the LiDAR coordinate system to the map coordinate system, and the updated points are added to the map. The final point cloud is as follows: Figure 6 shown.

[0091] B: Experimental analysis of laser radar detection of inspection robots;

[0092] 1) Dataset Introduction

[0093] We evaluate our method on the KITTI object detection dataset. The Kitti dataset consists of 7,481 training samples and 7,518 test samples. We further split the training samples into a training set of 3,712 samples and a validation set of 3,769 samples. Three difficulty levels for each class of objects are evaluated, and the main official evaluation metric is the average precision (AP) at a recall threshold of 40 (R40). Meanwhile, in the IoU metric, the IoU thresholds for cars, pedestrians, and cyclists are 0.7, 0.5, and 0.5, respectively.

[0094] 2) Training strategies and experimental platforms;

[0095] During the training process, we adopted widely used data augmentation strategies, including random scene flipping, random rotation, random scene scaling, and random translation. Our deep learning framework is Pytorch1.11.0, CUDA version is 11.3, batch_size is 4, epoch is 80, the starting learning rate is 0.003, and the optimizer is Adam.

[0096] 3) Quantitative analysis;

[0097] As shown in Table I, the detection results of our 3D detector and other algorithms in the BEV mode of the Kitti dataset are shown. All detection results are measured using the official KITTI evaluation detection indicators. The data results are all from the Kitti official benchmark or trained under the default configuration, and our model is calculated on a 3090 GPU. The KITTI dataset is divided into three levels of difficulty: easy, medium, and difficult. The official KITTI leaderboard is ranked according to the results of medium difficulty. As shown in Table I, in the prediction of vehicles in the difficult category, our method outperforms the baseline PV-RCNN with an AP (R40) of 5.78%. In particular, in the detection of small target cyclists, our algorithm outperforms the baseline with an AP (R40) of 9.69%, 4.05%, and 6.15%, respectively. In terms of pedestrian detection, it is 21.18%, 16.04%, 10.69%, and 5.51% higher than VoxelNet, SECOND, PointPillar, and Pillarnet, respectively. By comparing with other algorithms, it can be concluded that the improvement measures proposed by KASNet are effective and can significantly improve the accuracy of pedestrian detection while maintaining vehicle detection, which reflects the effectiveness of the algorithm.

[0098] TABLE I Bird’s-Eye View Test Performance: Results of the KITTI Test BEV Detection Benchmark

[0099]

[0100] C: Map update;

[0101] First, our inspection robot builds a 3D map of the unknown environment, and then converts the 3D point cloud map into a 2D map to achieve the navigation task. Secondly, during navigation, the three types of targets (cars, cyclists, and pedestrians) detected by the lidar are updated to the 2D map in the navigation to perform avoidance actions.

[0102] like Figure 7 (a) is the global 3D point cloud image detected by the lidar when there are pedestrians. Figure 7 (b) is a partial zoom of the 3D global map. The blue detection box marks the pedestrian point cloud area. In this experiment, in order to verify the real-time update of the map, we show the detection target and the non-detection target respectively, and show them through the final 2D navigation MAP. Figure 8 (a) is the 2D map without obstacles. Figure 8 (b) is the updated 2D map when there are pedestrians.

[0103] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A laser radar-based patrol inspection robot mapping navigation motion control method, characterized in that: include: Real-time acquisition of IMU data during the movement of the patrol inspection robot and 3D point cloud data obtained through 3D laser radar scanning; The 3D point cloud data is closed-loop fused with IMU data and the backward calculation process, and the radar sampling is compensated to build the map to obtain a 3D point cloud map. Detect obstacles encountered during movement in real time based on 3D point cloud data, and update the obstacles to the 2D map converted from the 3D point cloud map; The patrol detection robot is controlled to avoid obstacles based on the updated 2D map.

2. The laser radar-based patrol detection robot mapping navigation motion control method according to claim 1 is characterized in that: The 3D point cloud data is closed-loop integrated with the patrol detection robot’s IMU data and the backward calculation process, and the radar sampling and mapping are compensated to obtain a 3D point cloud map; specifically, it includes: Extract features of edge points and plane points in 3D point cloud data; Back-propagation motion compensation of 3D point cloud data based on the extracted features and the output of the forward propagation operation on the IMU data; The output of the backward motion compensation and the output of the forward operation are subjected to residual calculation and state update, and then a judgment is made. If it is judged to be a new observation point, it is updated to the 3D point cloud map.

3. The laser radar-based patrol detection robot mapping navigation motion control method according to claim 2 is characterized in that: Backward-forward motion compensation is performed based on the extracted features and the output of the forward-forward operation on the IMU data; specifically: In 3D LiDAR k At the sampling time, for multiple feature points between two adjacent frames of IMU data, the left IMU frame is used as the reference, and then the backward calculation is performed to project the relative pose of the point obtained by the backward calculation to time t k time, and then perform coordinate transformation.

4. The laser radar-based patrol detection robot mapping navigation motion control method according to claim 1 is characterized in that: Obstacles encountered during movement are detected in real time based on 3D point cloud data, and the obstacles are updated to the 3D point cloud map, including: Dynamic sparse coding branches are used to dynamically encode 3D point cloud data, and multi-scale sampling is used to learn multi-scale features of point clouds. Auxiliary supervision branches are used to alleviate the problem of spatial information blur and loss caused by downsampling of dynamic sparse coding branches; Use the key point branch to sample the key points of multi-scale features to obtain key point features; The ROI-grid pooling module is used to refine the proposals of multi-scale features and key point features, output the prediction results, and update the predicted obstacles to the 3D point cloud map.

5. The laser radar-based patrol detection robot mapping navigation motion control method according to claim 4 is characterized in that: The dynamic sparse coding branch consists of four sparse convolution blocks connected in sequence, which sample the 3D point cloud data by 1, 2, 4, and 8 times to obtain multi-scale features; The auxiliary supervision branch includes three upsampling blocks connected in sequence, corresponding to the last three sparse convolution blocks. The upsampling block first converts the voxel coordinates of the corresponding sparse convolution block into point coordinates in the real scene, then interpolates the sampling points through the feature propagation layer, and finally connects all point features to learn the structural knowledge characteristics of the target object; The key point branch includes four voxel set abstract blocks connected in sequence. The voxels of the four sparse convolution blocks are sampled by the farthest sampling method to select multiple key points. Each key point encodes the features of its adjacent non-empty voxels, abstracts the global and local features of the set and summarizes them into a key point feature.

6. A laser radar-based patrol inspection robot mapping and navigation device, characterized in that: include: The acquisition module is used to obtain the IMU data of the patrol detection robot during its movement and the 3D point cloud data obtained by 3D laser radar scanning in real time; Mapping module, which is used to close the loop of 3D point cloud data, fuse IMU data and backward calculation process, compensate radar sampling and map, and obtain 3D point cloud map; The obstacle detection and annotation module is used to detect obstacles encountered during movement in real time based on 3D point cloud data, and update the obstacles to the 2D map converted from the 3D point cloud map; The motion obstacle avoidance control module is used to perform motion obstacle avoidance control on the patrol detection robot based on the updated 2D map.

7. A patrol detection robot, characterized in that: It is configured to use the laser radar-based patrol detection robot mapping navigation motion control method described in any one of claims 1-5 for mapping and navigation.

8. 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 laser radar-based patrol inspection robot mapping navigation motion control method described in any one of claims 1 to 5 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the laser radar-based patrol inspection robot mapping navigation motion control method described in any one of claims 1 to 5 are implemented.

Images