An amr precise positioning and obstacle identification method of visual laser radar-free

Abstract

This invention provides a method for precise localization and obstacle recognition of AMRs without visual LiDAR, belonging to the field of LiDAR positioning technology. The method includes: constructing a lightweight pure topology map; acquiring the current frame point cloud, removing dynamic points from the current frame point cloud to obtain a static observation point cloud; extracting geometric features from the static observation point cloud and matching them with the lightweight pure topology map to obtain an initial pose estimate; performing scene segmentation based on the feature sparsity of the static observation point cloud and obtaining scene segmentation results; performing scene adaptive multi-scale matching to iteratively optimize the initial pose estimate to obtain the optimal global pose; performing hierarchical obstacle recognition on the current frame point cloud to determine the type of obstacle, and determining the spatial position of each obstacle based on the optimal global position; and matching an obstacle avoidance strategy based on the obstacle type and spatial position. This invention fundamentally solves the problem of AMR localization and obstacle avoidance in nighttime and low-light environments.

Description

Technical Field

[0001] This invention relates to the field of laser positioning technology, and more specifically to an AMR (Automatic Mobile Detection and Recognition) method using a visionless laser radar. Background Technology

[0002] With the development of smart logistics and intelligent manufacturing, AMRs (Autonomous Mobile Robots) have evolved from daytime auxiliary tools into 24 / 7 productivity equipment. Nighttime operations can complete tasks such as cargo handling, production line replenishment, and inventory counting during off-peak hours, significantly improving equipment utilization and site turnover efficiency.

[0003] The current mainstream AMR positioning and navigation technologies mainly include the following categories: (a) Pure Visual SLAM Solution This approach relies on monocular or binocular cameras to acquire environmental images, and then performs inter-frame matching and map construction by extracting feature points (such as ORB and SIFT features) from the images. While this approach has low hardware costs, it is heavily dependent on ambient lighting conditions.

[0004] (II) Vision-Liberation Fusion SLAM Solution This solution integrates both LiDAR and cameras, achieving localization and mapping through multi-sensor fusion (such as LIO-SAM, LVI-SAM frameworks). It performs exceptionally well in good daylight conditions, but the system as a whole relies on the normal operation of the visual modality.

[0005] (III) Traditional LiDAR SLAM Solution This scheme relies solely on lidar point cloud data for geometric registration and localization (e.g., LOAM, LeGO-LOAM algorithms). Localization accuracy is unaffected by lighting conditions, but obstacle recognition capabilities are limited, and the detection rate for low-reflectivity objects is low.

[0006] (iv) Magnetic track / QR code navigation solution This solution relies on pre-laid magnetic strips, magnetic nails, or QR code markers for path following, and is unaffected by light. However, it has high deployment costs, fixed paths, and cannot adapt to dynamically changing warehousing scenarios.

[0007] The existing technology has the following drawbacks: 1. Visual sensor performance drops sharply in low light environment: Whether it is pure visual SLAM or visual-laser fusion SLAM, the image signal-to-noise ratio of the camera module drops sharply in nighttime (the illuminance in warehouse environment is often below 10 lux, or even completely dark) environment, the feature extraction failure rate is high, resulting in frequent loss of localization or overall system degradation.

[0008] 2. Existing supplementary lighting solutions cannot fundamentally solve the problem: Some solutions attempt to improve nighttime imaging quality through infrared or LED supplementary lighting, but supplementary lighting will generate new interference factors such as reflection, shadows, and halos, which cannot guarantee system reliability and increases system power consumption and hardware costs.

[0009] 3. Traditional pure laser SLAM solutions are not good at recognizing low reflectivity objects: Most pure laser SLAM algorithms rely solely on geometric registration to achieve localization, and lack robust recognition capabilities for low reflectivity objects (such as black shelves with a reflectivity of about 5% and dark pedestrians with a reflectivity of about 15%), leading to missed detection or misjudgment of obstacles at night.

[0010] 4. Existing technologies lack systematic optimization for nighttime scenes: Most existing technical solutions are designed for general scenarios and have not been systematically adapted from hardware to algorithms for low-light environments at night, resulting in significant differences in daytime and nighttime performance.

[0011] 5. Large map data volume, difficult relocation at night: Existing visual maps or visual-laser fusion maps contain data such as image feature points, descriptors, and bag-of-words models, which require a large amount of storage space (200 MB to 2 GB). Moreover, in dark environments, due to the lack of visual references, the relocation success rate is low and the time consumption is long. Summary of the Invention

[0012] The purpose of this invention is to provide an AMR (Automatic Mapping) precise positioning and obstacle recognition method without visual LiDAR, so as to at least solve the problems of the prior art, such as the sharp decline in the performance of visual sensors in low light environment, lack of systematic optimization for night scene, and difficulty in night repositioning due to large map data volume.

[0013] To achieve the above objectives, the present invention provides a method for precise positioning and obstacle recognition using AMR (Automatic Mobile Detection) with a vision-free lidar, the method comprising: Extract point cloud data of the target scene, extract multi-level rigid geometric features that are inherent and structurally constant in the target scene from the point cloud data, and construct a lightweight pure topology map based on the multi-level rigid geometric features. Obtain the point cloud of the current frame, remove dynamic points in the point cloud of the current frame based on the temporal difference dynamic filtering rules, and obtain the static observation point cloud; Geometric features are extracted from static observation point clouds and matched with a lightweight pure topological map to obtain an initial pose estimate; Based on the feature sparsity of the static observation points, the target scene is segmented according to the feature sparsity to obtain the scene segmentation result; Based on the scene segmentation results, scene adaptive multi-scale matching is performed to iteratively optimize the initial pose estimation and obtain the optimal global pose. Perform hierarchical obstacle recognition on the current frame point cloud, determine the type of obstacle, and determine the spatial position of each obstacle based on the optimal global position; match obstacle avoidance strategies based on the type and spatial position of the obstacle.

[0014] Preferably, the time-series differential dynamic filtering rule is as follows: Based on a preset sliding window, extract historical point clouds from multiple frames that are adjacent to and continuous with the point cloud of the current frame. By unifying multiple historical point clouds into the same coordinate system as the current frame point cloud, the point cloud temporal sequence is obtained. Calculate the point cloud difference between each two adjacent frames of point cloud in the point cloud time sequence, and determine whether the point cloud difference is within the preset difference threshold. If yes, mark the corresponding point cloud as an overlapping point and remove the overlapping point. If no, mark the corresponding point cloud as a non-overlapping point. Track the spatial coordinate changes of each non-overlapping point within a preset sliding window to obtain the displacement change. When the displacement change continues to exceed the preset fluctuation threshold within the sliding window, the non-overlapping point is marked as a dynamic point. Remove all dynamic points from the current frame point cloud to obtain a static observation point cloud.

[0015] Preferably, the extraction of inherent and structurally constant multi-level rigid geometric features from point cloud data in the target scene includes: Calculate the local curvature of each point in the point cloud data, and perform classification on each point based on the local curvature to obtain the first level of global structured rigid features; Based on the multi-degree voxel filtering algorithm, the density gradient of point cloud data is calculated in a local region, the location of curvature abrupt change is identified, and the second-level local micro-geometric abrupt change features are obtained. Extract the height jump characteristics of point cloud data, estimate the ground height based on the height jump characteristics, and perform grid-by-grid region ground fitting based on the ground height to obtain the fitted ground; The boundary information of the fitted ground is extracted to obtain the third-level ground boundary features; the first-level global structured rigid features, the second-level local micro-geometric mutation features, and the third-level ground boundary features are used as the inherent and structurally constant multi-level rigid geometric features in the target scene.

[0016] Preferably, the nodes of the lightweight pure topology map are used to store node location coordinates, node type, multi-level rigid geometric feature descriptors, and node ID and timestamp; wherein, the node type includes corridor nodes, intersection nodes, work nodes, and landmark nodes; The edges of the lightweight pure topology map are used to store the relative pose transformations between nodes, edge types, and information matrices. The information in the lightweight pure topology map includes the map version number, construction time, coordinate system definition, and feature sparsity distribution.

[0017] Preferably, the scene segmentation of the target scene is performed based on the feature sparsity of the static observation points, including: Count the number of valid feature points and the total number of points in the static observation point cloud; Feature sparsity is generated based on the ratio of effective feature points to total number of points; When the feature sparsity is within a preset threshold range, the target scene is divided into a regular region; When the feature sparsity is greater than the maximum value of the threshold region, the target scene is divided into a task docking area; When the feature sparsity is less than the minimum value of the threshold region, the target scene is divided into an open passage area.

[0018] Preferably, scene adaptive multi-scale matching is performed based on the scene segmentation results, including: When the scene segmentation result is an open passage area, the initial pose estimation is iteratively optimized based on a coarse-scale global topology matching strategy to obtain the optimal global pose. When the scene segmentation result is the work docking area, the initial pose estimation is iteratively optimized based on the fine-scale local micro-curvature matching strategy to obtain the optimal global pose. When the scene segmentation result is a regular region, the initial pose estimation is iteratively optimized by a fusion strategy based on coarse-scale global topology matching and fine-scale local micro-curvature matching to obtain the optimal global pose.

[0019] Preferably, the coarse-scale global topology matching strategy is as follows: Downsampling is performed on the static observation point cloud to generate a sparse feature point cloud; A target grid probability model is constructed based on a lightweight pure topological map, wherein each grid of the target grid probability model pre-stores the mean vector and covariance matrix of the point cloud; The sparse feature point cloud is projected onto the map coordinate system according to the current transformation parameters to obtain several projection points. The initial value of the current transformation parameters is the initial pose estimate. Based on the target grid probability model, calculate the sum of the probability densities of each projection point in each grid. The current transformation parameters are updated using the Jacobian matrix to maximize the sum of probability densities; Extract the poses of two consecutive iterations. Stop iterating when the change in pose between two iterations reaches a preset threshold to obtain the final pose. Use the final pose as the optimal global pose.

[0020] Preferably, the fine-scale local micro-curvature matching strategy is as follows: Extract acute-angle feature points and planar feature points from the static observation point cloud as the source point cloud, and perform downsampling on the source point cloud to generate a sampled source point cloud; Extract subacute angle feature points and subplane feature points near the current pose from a lightweight pure topology map to generate a target point cloud, and perform downsampling on the target point cloud to generate a sampled target point cloud; where the initial value of the current pose is the initial pose estimate; Based on the KD-tree nearest neighbor search algorithm, the nearest neighbor point is obtained by searching the target point cloud and the nearest neighbor point of each point in the source point cloud. An error function is constructed based on the nearest neighbor point, the sampled source point cloud, and the rotation matrix and translation vector to be solved. The Levenberg-Marquardt algorithm is used to solve for the rotation matrix and translation vector to obtain the optimal rotation matrix and translation vector. The optimal global pose is then determined based on the optimal rotation matrix and translation vector.

[0021] Preferably, the fusion strategy for coarse-scale global topology matching and fine-scale local micro-curvature matching is as follows: A coarse-scale global topology matching strategy is executed at low frequency to obtain the first pose estimate and the corresponding confidence score. The second pose estimate is obtained by performing a fine-scale local micro-curvature matching strategy at high frequency. When the confidence score is greater than the confidence threshold, the first pose estimate is used as the initial value of the current pose of the fine-scale local micro-curvature matching strategy, and the second pose estimate is used as the optimal global pose. When the confidence score is greater than the confidence threshold, a fusion weight is generated based on a nonlinear least squares function; the first pose estimate and the second pose estimate are weighted and fused based on the fusion weight to obtain the optimal global pose.

[0022] Preferably, the lidar of the AMR includes a main lidar and two auxiliary lidars. The main lidar is a three-dimensional multi-threaded lidar, and the auxiliary lidars are two-dimensional single-threaded lidars. The main lidar is used to acquire a three-dimensional point cloud around the AMR, and the two auxiliary lidars are used to acquire a two-dimensional point cloud positioned above and below the front of the AMR, respectively. The current frame point cloud is a three-dimensional point cloud around the AMR. Perform hierarchical obstacle recognition on the current frame point cloud, including: Ground separation is performed on the two-dimensional point cloud located in front of the AMR based on the CSF algorithm to obtain the ground point cloud; Negative obstacle detection is performed on the ground point cloud to obtain the ground negative obstacle recognition results; High-altitude suspended object detection is performed based on the two-dimensional point cloud positioned in front of the AMR to obtain high-altitude obstacle recognition results; Based on the ground point cloud, remove the ground points from the current frame point cloud to obtain the current frame point cloud after removal; Density clustering is performed on the point cloud of the current frame after the removal of obstacles to obtain multiple independent point cloud clusters; each point cloud cluster represents a potential obstacle. Geometric parameters are calculated for each point cloud cluster to generate three-dimensional geometric indices for the point cloud clusters. Based on the three-dimensional geometric indices, ground positive obstacle identification results are determined. These ground positive obstacle identification results are used to update the lightweight pure topology map. The number of point clouds in a single frame within a point cloud cluster is counted. If the number of point clouds in a single frame is less than a preset number, the point cloud cluster is marked as a suspected low-reflectivity object. Based on the multi-frame temporal accumulation rule, density accumulation is performed on the point cloud clusters suspected of being low reflectivity objects. Then, the point cloud clusters after density accumulation are subjected to point cloud cluster reflection intensity and spatial density analysis and stability verification to obtain the low reflectivity object recognition results.

[0023] The beneficial effects of this invention are: 1. This invention fundamentally solves the problem of positioning and obstacle avoidance of AMRs in nighttime and low-light environments by constructing a completely vision-free pure laser perception architecture and matching it with multi-level rigid geometric feature extraction, temporal differential dynamic filtering, scene adaptive matching, and layered obstacle recognition algorithms.

[0024] 2. This invention utilizes multi-level rigid geometric features to construct a lightweight pure topological map, so that the map only stores data such as topological relationships, curvature, angles and size parameters related to the multi-level rigid geometric features, and does not store any image, texture or color information; the lightweight pure topological map of the same scale is much smaller than the visual map.

[0025] 3. The lightweight pure topological map construction method of this invention produces maps that are identical to those built during the day, eliminating positioning errors caused by differences between day and night maps. There is no need to build two separate day and night maps for the same location. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the AMR precise positioning and obstacle recognition method using a vision-free lidar provided by the present invention. Detailed Implementation

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0028] Example 1 The system architecture of the AMR (Autonomous Mobile Robot) precise localization and obstacle recognition method without vision LiDAR in this embodiment includes: an AMR body and multiple sensors, excluding vision sensors. The multiple sensors include: A multi-line 3D master LiDAR (such as a 16-line / 32-line / 64-line LiDAR with a frequency of 10 Hz): responsible for main positioning and 3D environmental modeling, installed on the top of the AMR body, with a scanning range covering a 360° horizontal field of view and ±15° vertical field of view; At least two two-dimensional auxiliary lidar units (single-line, frequency 20 Hz~50 Hz): deployed at the upper and lower ends of the front of the vehicle body respectively: the upper lidar is responsible for scanning high-altitude suspended objects (suspended signs, crossbeams, pipes), the upper lidar is 2.0 m~2.5 m above the ground, and the scanning plane is at an elevation angle of 15°~30° with the horizontal plane; the lower lidar is responsible for detecting negative obstacles such as potholes, depressions, and curbs on the ground; the lower lidar is 0.2 m~0.5 m above the ground, and the scanning plane is at a depression angle of 15°~30° with the horizontal plane.

[0029] Inertial Measurement Unit (IMU): Installed near the geometric center of the vehicle body and strictly aligned with the vehicle coordinate system, it provides high-frequency (100~200 Hz) angular velocity and acceleration information to compensate for point cloud distortion caused by vehicle movement during lidar scanning; Odometer: Provides vehicle speed constraints and assists in short-term positioning.

[0030] The coordinate system calibration of the multi-laser radar of this invention is as follows: Spatial geometric structural features are used to group overlapping point cloud data from multiple lidar systems, and linear features are selected to achieve coordinate system unification. Specifically, during the calibration phase, calibration objects with known geometric features (such as planar plates or right-angle prisms) are placed within the common field of view of the multiple lidars. By extracting the same geometric features from the point clouds of each lidar system, a coordinate system transformation matrix T={R,t} is established. r}, where R is the rotation matrix and t r It serves as a translation vector to achieve spatial alignment of multi-radar point cloud data.

[0031] An optimization framework based on tightly coupled multi-source information is adopted, incorporating multi-sensor information as constraints into the system for joint optimization. The system includes the following constraint types: 1. IMU pre-integration constraint: Pre-integrate the IMU data between two frames to provide relative motion constraints. At the same time, the estimated motion information is used to eliminate point cloud distortion and provide initial values ​​for LiDAR odometry optimization. 2. LiDAR odometry constraints: Constraints are constructed by matching frames with local maps; 3. Loop closure detection constraint: When the robot returns to the visited area, a global constraint is constructed through loop closure detection to eliminate accumulated errors; Multi-machine collaborative constraints (in multi-machine deployment): Multiple AMRs exchange point cloud data and pose information, share loop closure detection constraints, and optimize collective location. Figure 1 To the point of being responsive.

[0032] Based on the above framework, such as Figure 1 As shown, this embodiment provides a method for precise AMR positioning and obstacle recognition using a vision-free LiDAR, the method comprising: Step S10: Extract point cloud data of the target scene, extract the inherent and structurally constant multi-level rigid geometric features in the target scene from the point cloud data, and construct a lightweight pure topology map based on the multi-level rigid geometric features.

[0033] In this embodiment of the invention, the target scenarios include, but are not limited to: nighttime off-peak picking scenarios in e-commerce warehouses, night shift material delivery scenarios in automobile assembly workshops, and unmanned inspection scenarios in underground mines.

[0034] In this embodiment of the invention, the lightweight pure topology map is constructed using point cloud data collected by the main lidar. During the initialization phase, when the AMR starts from a standstill, the system directly defines the current vehicle pose as the origin (0, 0, 0) of the global coordinate system. The attitude angles roll (roll angle, the angle of rotation around the front-to-back axis), pitch (pitch angle, the angle of rotation around the side-to-side axis), and yaw (yaw angle, the angle of rotation around the vertical axis) are initialized by the gravity direction measured by the IMU in a standstill state (where roll and pitch are calculated by gravitational acceleration, and yaw is set to 0 by default or given by the magnetometer / preset heading).

[0035] Then, the first frame of point cloud data acquired by the main LiDAR is used as the point cloud data of the target scene to create a lightweight pure topology map.

[0036] In this embodiment of the invention, extracting multi-level rigid geometric features inherent and structurally constant in the target scene from point cloud data includes: Step S101: Calculate the local curvature of each point in the point cloud data, and perform classification on each point based on the local curvature to obtain the first-level global structured rigid features.

[0037] In this embodiment of the invention, the first level is mainly used to extract large-scale structures such as walls, columns, and shelf outlines in the environment.

[0038] The formula for calculating the local curvature in this embodiment is as follows:

[0039] In the formula, c is the local curvature of point X in the point cloud data, and N is the total number of points. Let X be the distance from point X to the origin of the lidar. Let X be the i-th point in the neighborhood of point X. Let X be a point and each point X in its neighborhood. i The sum of the difference vectors; this formula measures the smoothness of a local point cloud: the curvature of points on a straight line is 0, while the curvature of sharp edges or inflection points is larger.

[0040] The classification strategy in this embodiment is as follows: 1. Divide each laser scanning line into 6 equal parts; 2. The blocks are arranged according to the magnitude of curvature; 3. The two points with the greatest curvature are marked as corner_sharp (acute angle feature points); 4. The 20 points with the largest curvature are marked as corner_less_sharp (secondary acute angle feature points); 5. The four points with the smallest curvature are marked as surf_flat (planar feature points); 6. Other points and surf_flat→surf_less_flat (subplane feature points); In this embodiment of the invention, the following matching relationship exists between two adjacent point clouds: the corner_sharp point of the current frame matches the corner_less_sharp point of the previous frame; the surf_flat point of the current frame matches the surf_less_flat point of the previous frame.

[0041] This embodiment establishes a sparse but highly reliable point-to-point / point-to-surface association between two adjacent frames, thereby quickly and robustly calculating the robot's relative motion increment within a very short time interval.

[0042] Step S102: Based on the multi-degree voxel filtering algorithm, calculate the density gradient of the point cloud data in the local area, identify the curvature change location, and obtain the second-level local micro-geometric change features; the second-level local micro-geometric change features in this embodiment are small-scale geometric changes such as equipment edges, docking holes, and the roots of shelf supports.

[0043] Step S103: Extract the height jump characteristics of the point cloud data, estimate the ground height based on the height jump characteristics, perform grid-by-grid region ground fitting based on the ground height to obtain the fitted ground; extract the boundary information of the fitted ground to obtain the third-level ground boundary features; the third-level ground boundary features include features such as curbs, pothole edges, and ground joints.

[0044] The first-level global structured rigid features, the second-level local micro-geometric mutation features, and the third-level ground boundary features are used as the inherent and structurally constant multi-level rigid geometric features in the target scene.

[0045] Therefore, lightweight pure topological maps are created using multi-level rigid geometric features, all derived from physical geometry and completely independent of surface color, texture, and lighting conditions. In complete darkness, the curvature, angles, and dimensions of these geometric features remain identical to those during the day. This ensures that nighttime and daytime maps are completely identical, eliminating positioning errors caused by differences between day and night maps. There is no need to construct separate day and night maps for the same site.

[0046] The lightweight pure topology map in this embodiment includes node, edge, and primitive information; The nodes are used to store the node location coordinates (x, y, z), node type (corridor node, intersection node, work node, and landmark node), multi-level rigid geometric feature descriptors (feature point curvature distribution histogram, plane normal vector), and node ID and timestamp. The edges of the lightweight pure topology map are used to store the relative pose transformations between nodes (translation vector + rotation matrix / quaternion), edge types (odometry constraint / scan matching constraint / loop closure detection constraint), and information matrix (representing the uncertainty of the constraints). The information in the lightweight pure topology map includes the map version number, construction time, coordinate system definition, and feature sparsity distribution.

[0047] Step S20: Obtain the current frame point cloud, remove dynamic points in the current frame point cloud based on the temporal differential dynamic filtering rule, and obtain the static observation point cloud; wherein, the current frame point cloud is the point cloud data of the second frame and subsequent frames.

[0048] In this embodiment of the invention, dynamic objects (such as forklifts and people) in a nighttime environment have blurred outlines and are difficult to distinguish visually. In order to accurately identify and eliminate moving objects and avoid dynamic points from contaminating the positioning results, this embodiment uses a temporal differential dynamic filtering rule to accurately identify and eliminate moving objects.

[0049] Specifically, the time-series differential dynamic filtering rule is as follows: Step S201: Based on a preset sliding window, extract multiple historical point clouds that are adjacent to and continuous with the current frame point cloud; the number of historical point clouds is usually 5 to 10 frames.

[0050] Step S202: Unify the historical point clouds of multiple frames into the same coordinate system as the point cloud of the current frame to obtain the point cloud time sequence.

[0051] In this embodiment of the invention, since the robot is in motion, even stationary objects have different coordinates in different frames. Before differential processing, the point clouds of historical frames need to be unified to the coordinate system of the current frame. Specifically, the point cloud of the previous frame is projected onto the coordinate system of the current frame using a known pose transformation to obtain an aligned point cloud. Multiple aligned point clouds are then sequentially ordered in time to form a point cloud temporal sequence.

[0052] Step S203: Calculate the point cloud difference between each two adjacent point clouds in the point cloud time sequence, and determine whether the point cloud difference is within the preset difference threshold. If yes, mark the corresponding point cloud as an overlapping point and remove the overlapping point. If no, mark the corresponding point cloud as a non-overlapping point.

[0053] The formula for calculating the point cloud difference in this embodiment is as follows:

[0054] In the formula, Let be the point cloud difference at time t. For the point cloud at time t in the point cloud time series, This is a set difference operation used to remove spatial overlap between two frames.

[0055] Step S204: Track the spatial coordinate changes of each non-overlapping point within a preset sliding window to obtain the displacement change; wherein, the sliding window is a window of 3 consecutive frames.

[0056] Step S205: When the displacement change is continuously greater than the preset fluctuation threshold within the sliding window, the non-overlapping point is marked as a dynamic point; wherein, the fluctuation threshold is the value of the fluctuation threshold; in this embodiment, when the displacement change is less than the preset fluctuation threshold within the sliding window, or exhibits random walk and irregular jumping, it is regarded as noise and ignored.

[0057] Step S206: Remove all dynamic points in the current frame point cloud to obtain a static observation point cloud.

[0058] This embodiment can avoid the dynamic points from contaminating the positioning results by accurately identifying dynamic points, thereby achieving dynamic interference suppression.

[0059] In this embodiment of the invention, the oldest frame in the sliding window is discarded, the point cloud of the current frame is pushed into the end of the window, and the latest optimal global pose calculated by step S50 is used as the initial value for motion compensation in the next frame.

[0060] Step S30: Extract geometric features from the static observation point cloud and match them with a lightweight pure topology map to obtain an initial pose estimate; specifically: match the geometric features (corner_sharp / surf_flat) extracted in the current frame with the features (corner_less_sharp / surf_less_flat) in the lightweight topology map to obtain an initial pose estimate.

[0061] Step S40: Based on the feature sparsity of the static observation points, perform scene segmentation of the target scene according to the feature sparsity to obtain the scene segmentation result.

[0062] In this embodiment of the invention, pure lidar inherently presents a contradiction in its physical characteristics: "accurate ranging but semantic blind spots." A single algorithm cannot simultaneously achieve both "global accuracy" and "local precision" in any environment. For example, the laser matching algorithm relies on the geometric features of the environment (lines at corners, surfaces on the ground) to constrain the degrees of freedom. However, in open passages at night or large warehouse corridors, the environment resembles a "long, straight tunnel," with extremely sparse point cloud features, resulting in a "geometric degradation" dilemma. Furthermore, AMRs not only have the need for mobility but also the rigid requirement for precise docking (such as diving into shelves and retrieving materials).

[0063] In response, this embodiment divides the target scene into segments and performs subsequent scene adaptive multi-scale matching, enabling the AMR to have "macroscopic memory" in open areas so as not to get lost, and "microscopic resolution" in the work area so as not to collide with other vehicles, ultimately achieving highly available autonomous navigation in all weather and all terrain.

[0064] Specifically, based on the feature sparsity of the static observation point cloud, the target scene is divided according to the feature sparsity, including: counting the number of effective feature points and the total number of points in the static observation point cloud; generating feature sparsity based on the ratio of the number of effective feature points to the total number of points; dividing the target scene into a regular area when the feature sparsity is within a preset threshold area; dividing the target scene into a work docking area when the feature sparsity is greater than the maximum value of the threshold area; and dividing the target scene into an open passage area when the feature sparsity is less than the minimum value of the threshold area.

[0065] In this embodiment, the effective features are the corner points and planar feature points in the point cloud, and the formula for calculating feature sparsity is:

[0066] In the formula, For characteristic sparsity, This represents the total number of corner points and planar feature points, i.e., the number of effective feature points. This represents the total number of points in the static observation point cloud.

[0067] In this embodiment, the threshold range is 0.05 to 0.20. That is, when S < 0.05, the target scene is divided into a normal area; when 0.05 ≤ S ≤ 0.20, the target scene is divided into a normal area; when S > 0.20, the target scene is divided into a work docking area.

[0068] Step S50: Perform scene adaptive multi-scale matching based on the scene segmentation results to perform iterative optimization on the initial pose estimation and obtain the optimal global pose.

[0069] In this embodiment, when the scene segmentation result is an empty passage area, it indicates that the environment is extremely open (such as a long passage at night) with very few feature points. At this time, the system abandons fine point-to-point matching and instead relies on a large-scale contour. That is, iterative optimization is performed on the initial pose estimation based on a coarse-scale global topology matching strategy to obtain the optimal global pose.

[0070] Specifically, the coarse-scale global topology matching strategy is as follows: Step A1: Downsampling the static observation point cloud to generate a sparse feature point cloud; specifically: voxel filtering is performed on the static observation point cloud at a resolution of 0.5 m to generate a sparse feature point cloud. Step A2: Construct a target grid probability model based on a lightweight pure topology map. Each grid in the target grid probability model stores the mean vector and covariance matrix of the point cloud. Specifically, the lightweight pure topology map is divided into grids with the same 0.5 m resolution, and each grid stores the mean vector and covariance matrix of the point cloud (i.e., the NDT probability distribution model).

[0071] Step A3: Project the sparse feature point cloud onto the map coordinate system according to the current transformation parameters to obtain several projection points. The initial value of the current transformation parameters is the initial pose estimate.

[0072] Step A4: Based on the target grid probability model, calculate the sum of the probability densities of each projection point in each grid.

[0073] Step A5: Update the current transformation parameters using the Jacobian matrix to maximize the sum of probability densities; the maximum number of iterations in this embodiment is 20.

[0074] Step A6: Extract the poses from two consecutive iterations. Stop iterating after the change in pose between two iterations reaches a preset threshold, obtaining the final pose. Use this final pose as the optimal global pose. The preset threshold is a pose change of less than 1 × 10⁻⁶. -4 (Translation is in meters, rotation is in rads); the optimal global pose is a 6-DOF pose transformation (x, y, z, roll, pitch, yaw).

[0075] The coarse-scale global topology matching strategy in this embodiment is suitable for maintaining localization in scenarios with sparse geometric features, relying on a large-scale topological contour to ensure that localization is not lost.

[0076] For environments with dense features (such as shelves and docking holes), the system strives for millimeter-level high precision; that is, when the scene is divided into a work docking area, the initial pose estimation is iteratively optimized based on a fine-scale local micro-curvature matching strategy to obtain the optimal global pose.

[0077] Specifically, the fine-scale local micro-curvature matching strategy is as follows: Step B1: Extract acute-angle feature points and planar feature points from the static observation point cloud as the source point cloud, and perform downsampling on the source point cloud to generate a sampled source point cloud; Step B2: Extract subacute angle feature points and subplanar feature points (approximately 200-500 points in total) from the lightweight pure topology map near the current pose to generate the target point cloud. Then, perform downsampling on the target point cloud to ensure that the number of points does not exceed 300 points in order to control the computation time and generate the sampled target point cloud. The initial value of the current pose is the initial pose estimate.

[0078] Step B3: Based on the KD-tree nearest neighbor search algorithm, search for the nearest neighbor point in the sampled target point cloud to each point in the sampled source point cloud to obtain the nearest neighbor point; wherein, the search radius of the KD-tree nearest neighbor search algorithm is initially set to 0.5 m, and gradually reduced to 0.1 m with iteration.

[0079] Step B4: Construct an error function based on the nearest neighbor point, the sampled source point cloud, and the rotation matrix and translation vector to be solved; the formula for calculating the error function is:

[0080] In the formula, e is the error function. For the points of the sampled source point cloud, The target point cloud is the nearest neighbor of each point in the source point cloud. for The normal vector at that point, For rotation matrix, It is a translation vector.

[0081] Step B5: Solve the rotation matrix and translation vector based on the Levenberg-Marquardt algorithm to obtain the optimal rotation matrix and translation vector, and determine the optimal global pose based on the optimal rotation matrix and translation vector.

[0082] The maximum number of iterations is 30, and the convergence condition is that the change in the mean square error between two consecutive iterations is less than 1 × 10⁻⁶. -6 m 2 Or the rotational change is less than 1×10 -5 rad and the translation change is less than 1×10 -5 m. The output is then a 6-DOF pose transformation and its covariance matrix.

[0083] In this embodiment, when the AMR enters the docking area, it needs to meet the rigid requirement of precise docking (such as diving into the shelf and inserting materials). However, in order to pursue global convergence, the coarse-scale global topology matching strategy adopts large-scale voxel filtering (0.5m resolution), which causes it to naturally lose fine structure. The ultimate accuracy can only reach the decimeter level (±5cm), which is completely unable to meet the docking accuracy requirement (±3mm).

[0084] Therefore, by introducing a fine-scale local micro-curvature matching strategy, this embodiment can seamlessly transition from "extensive topology navigation" to "fine micro-curvature matching", enabling millimeter-level fine-tuning of the AMR.

[0085] For regular regions, this embodiment employs a fusion strategy of coarse-scale global topology matching and fine-scale local micro-curvature matching to perform iterative optimization on the initial pose estimation to obtain the optimal global pose; that is, the system executes two threads in parallel: coarse-scale global matching and fine-scale local matching.

[0086] Specifically, the fusion strategy for coarse-scale global topology matching and fine-scale local micro-curvature matching is as follows: Step C1: Execute the coarse-scale global topology matching strategy at a low frequency to obtain the first pose estimate and the corresponding confidence score; wherein, the coarse-scale global topology matching strategy is run at a low frequency (1~2 Hz); the confidence score is the trace of the covariance matrix corresponding to the first pose estimate.

[0087] Step C2: Execute the fine-scale local micro-curvature matching strategy at a high frequency to obtain the second pose estimate; wherein, the fine-scale local micro-curvature matching strategy is run at a higher frequency (10 Hz).

[0088] Step C3: When the confidence score is greater than the confidence threshold, the first pose estimate is used as the initial value of the current pose of the fine-scale local micro-curvature matching strategy, and the second pose estimate is used as the optimal global pose.

[0089] In this embodiment, the confidence threshold is 0.7. When the confidence score is higher than the threshold of 0.7, the system uses the first pose estimate of the coarse matching output as the initial value input of the fine matching to accelerate the convergence of the fine matching and prevent it from getting trapped in local extrema.

[0090] Step C4: When the confidence score is greater than the confidence threshold, generate fusion weights based on the nonlinear least squares function; perform weighted fusion on the first pose estimate and the second pose estimate based on the fusion weights to obtain the optimal global pose.

[0091] When the confidence score is below 0.7, the system determines that the current scene has a high risk of localization degradation. In this case, the second pose estimate from the fine-matching is only included as a local constraint in the optimization framework, while the first pose estimate from the coarse-matching provides global prior constraints. The two are then fused using a weighted fusion of information matrices within a nonlinear least-squares optimization framework, ultimately outputting the fused optimal pose estimate.

[0092] The functional expression of the nonlinear least squares optimization framework in this embodiment is:

[0093] In the formula, For lightweight pure topology map nodes With nodes The residual function of the edges between them. This is an information matrix.

[0094] In the final step of the coarse-scale global matching strategy in this embodiment, the approximate Hessian matrix (H) actually reflects the curvature of the cost function at the optimal solution. The Hessian matrix constructs the information matrix of the coarse-scale global matching strategy. In the fine-scale local matching strategy, the Jacobian matrix and residual vector of all corresponding point pairs at the final iteration are extracted. Based on the residual vector and the Jacobian matrix, the covariance matrix can be calculated, and the inverse of the covariance matrix is ​​the information matrix of the fine-scale local matching strategy. Then, fusion weights are created based on the ratio between the trace of the information matrix of the coarse-scale global matching strategy and the trace of the information matrix of the fine-scale local matching strategy.

[0095] Next, a global prior residual factor (i.e., the Lie algebra difference between the current pose to be optimized and the coarse matching output) is constructed based on the first pose estimate, and a local high-frequency residual factor (the Lie algebra difference between the first pose estimate and the second pose estimate) is constructed based on the second pose estimate. The global prior residual factor and the local high-frequency residual factor are inserted into the nonlinear least squares optimization framework. Then, the information matrices of the coarse and fine scales in the nonlinear least squares optimization framework are updated using fusion weights (i.e., the fusion weights of the fine-scale local matching strategy are multiplied by the information matrix of the fine-scale local matching strategy to obtain the updated fine-scale information matrix). Finally, the nonlinear optimizer (Levenberg-Marquardt) is used to perform iterations to obtain the final optimal pose estimate.

[0096] The output of the scene adaptive multi-scale matching in this embodiment consists of two parts: (1) the 6-DOF pose transformation (x, y, z, roll, pitch, yaw) of the AMR at the current time, i.e., the localization result; (2) the information matrix (or covariance matrix) of the pose estimation, which is used to represent the confidence level of the localization result. The above output is directly used in two aspects: First, as an important constraint (LiDAR odometry constraint) in the aforementioned multi-source tightly coupled optimization framework, it participates in nonlinear least squares optimization together with the IMU pre-integration constraint and loop closure detection constraint to update the global pose estimation and map; Second, it provides an accurate projection reference of the current frame point cloud in the global coordinate system for the subsequent step S60, so that the spatial position determination of obstacles and map marking have global consistency.

[0097] Step S60: Perform hierarchical obstacle recognition on the current frame point cloud, determine the type of obstacle, and determine the spatial position of each obstacle based on the optimal global position; match obstacle avoidance strategies based on the type and spatial position of the obstacle.

[0098] In this embodiment of the invention, the lidar of the AMR includes a main lidar and two auxiliary lidars. The main lidar is a three-dimensional multi-threaded lidar, and the auxiliary lidars are two-dimensional single-threaded lidars. The main lidar is used to acquire a three-dimensional point cloud around the AMR, and the two auxiliary lidars are used to acquire a two-dimensional point cloud positioned above and below the front of the AMR, respectively. The current frame point cloud is a three-dimensional point cloud around the AMR.

[0099] Therefore, the specific steps for performing layered obstacle recognition on the current frame point cloud are as follows: Step S601: Perform ground separation on the two-dimensional point cloud placed in front of the AMR based on the CSF (Cloth Simulation Filtering) algorithm to obtain the ground point cloud.

[0100] Step S602: Perform negative obstacle detection on the ground point cloud to obtain the ground negative obstacle recognition results; ground negative obstacles include obstacles such as potholes and ditches, which are not affected by light at night, and the detection rate is the same as during the day.

[0101] Step S603: Perform high-altitude suspended object detection based on the two-dimensional point cloud in front of the AMR to obtain the high-altitude obstacle identification result; high-altitude obstacles include suspended signs, beams, pipes, drooping cables, etc.

[0102] Step S604: Based on the ground point cloud, remove the ground points in the current frame point cloud to obtain the removed current frame point cloud.

[0103] Step S605: Perform density clustering on the removed point cloud of the current frame to obtain multiple independent point cloud clusters; each point cloud cluster represents a potential obstacle.

[0104] Step S606: Perform geometric parameter calculations on each point cloud cluster to generate three-dimensional geometric indices for the point cloud clusters. Based on the three-dimensional geometric indices, determine the ground positive obstacle recognition results. The ground positive obstacle recognition results are used to update the lightweight pure topology map. The geometric parameters include height, horizontal projected area, and contour curvature.

[0105] Step S607: Count the number of point clouds in a single frame in the point cloud cluster. When the number of point clouds in a single frame is less than a preset number, the point cloud cluster is marked as a suspected low reflectivity object.

[0106] Step S608: Based on the multi-frame temporal accumulation rule, perform density accumulation on the point cloud clusters of suspected low reflectivity objects, and perform point cloud cluster reflection intensity and spatial density analysis and stability verification on the point cloud clusters after density accumulation to obtain the low reflectivity object recognition result.

[0107] In this embodiment, low-reflectivity objects (black goods with a reflectivity of approximately 5%, dark-colored pedestrians, and polished surfaces) have weak laser reflection, resulting in extremely sparse point clouds in a single frame. Existing technologies cannot effectively identify these objects. Therefore, this embodiment employs a multi-frame temporal accumulation rule: continuously accumulate M frames of point clouds (M=10~20 frames recommended), load the accumulated frames into the coordinate space of the current frame, and through accumulation and superposition, increase the density of the originally sparse target point cloud to above a recognizable threshold. The rate of change in point cloud density in the region after accumulation is statistically analyzed (to determine if it changes from sparse to dense), and the reflection intensity distribution is also analyzed (the echo intensity of low-reflectivity objects is significantly lower than that of the surrounding high-reflectivity environment). Furthermore, non-maximum suppression is applied to multi-frame observations at the same spatial location to remove occasionally appearing noise points, retaining only echo points that are consistently and stably present across multiple frames.

[0108] The comprehensive criteria for judging low reflectivity objects are: 1. Sparse in a single frame and significantly increased density after accumulation in multiple frames; 2. Stable spatial position in multiple consecutive frames (high temporal stability); 3. Physically reasonable geometric contours (such as the edge of a rectangular shelf or the contour of a human body).

[0109] The obstacle avoidance strategy in this embodiment is shown in Table 1: Table 1. Response Table for Hierarchical Obstacle Avoidance Strategy

[0110] As a further optimization of this embodiment, the method further includes: performing a self-check based on the health status of a lightweight pure topology map to trigger relocation.

[0111] Specifically, define a map health metric H, which includes: the inlier ratio of the current frame to the map; the average reprojection error between the current frame's feature points and map feature points; and the rate of change of pose covariance over the most recent W frames (W=100 recommended). If H is below a threshold τ... H (If the matching rate of in-points is less than 30%), a relocation process will be triggered.

[0112] The current frame point cloud is downsampled to a sparse point cloud (resolution 0.5 m) using voxel filtering. Then, all nodes in the topological map are traversed, and for each node, its stored local geometric feature descriptors (feature point curvature distribution histogram and plane normal vector) are extracted. The similarity score between the current frame sparse point cloud and the feature descriptors of each node is calculated. The scoring method uses a weighted sum of the angle between the histogram intersection kernel and the normal vector, with weights of 0.6 and 0.4, respectively. The top 5 nodes with the highest scores are selected as candidates. In the fine matching verification stage, for each candidate node, the original feature point cloud of the current frame is matched with the local map point cloud in the neighborhood of that node using fine-scale local micro-curvature matching (point-to-surface distance measurement). The maximum number of iterations is 30, and the convergence threshold is 1×10⁻⁶. -3 Finally, the candidate with the smallest reprojection error is selected as the relocalization result, and the information matrix of this pose is output as the initial confidence level for subsequent localization and tracking.

[0113] As a further optimization of this embodiment, the method further includes: performing collaborative calibration on multiple AMRs, wherein the multiple AMRs interact with point cloud data and pose information during operation, and collectively optimize the location by sharing loop closure detection constraints. Figure 1 Consistency is maintained, and the system remains effective even at night; cluster collaboration is unaffected by day or night. The specific steps are as follows: Each AMR continuously detects loop closure events (i.e., successful matching events between the current frame point cloud and its own historical map) during its operation. When an AMR detects a loop closure, it broadcasts the residual vector and information matrix of the loop closure constraint, along with the current global pose estimate, to other AMRs within the operating area via wireless communication. Upon receiving the loop closure information, other AMRs first perform a spatiotemporal consistency check—comparing the received loop closure pose with their current pose. If the relative pose deviation is less than a preset threshold (translation < 1.0 m, rotation < 15°), it is considered a valid cooperative constraint; otherwise, it is considered a potential data conflict, triggering cross-validation of the point cloud data between the two AMRs (i.e., projecting the current frame point cloud onto the other AMR's loop closure pose coordinate system for matching verification). Verified cooperative constraints are added as additional edges (edge ​​type: multi-machine cooperative constraint) to the global optimization graph of each AMR, participating in subsequent nonlinear least squares optimization. The weights of the information matrix of the aforementioned cooperative constraints are jointly determined by the covariance matrices of the pose estimates of both AMRs, with a weight coefficient of W. coop = (Σ1 -1 + Σ2 -1 ) / 2, where Σ1 and Σ2 are the pose covariance matrices of the two machines at the time of cooperation.

[0114] The collaborative calibration in this embodiment ensures that multiple AMRs can maintain a consistent map representation within a shared operating area, and the cumulative positioning error of a single machine is effectively suppressed through collaborative constraints.

[0115] Example 2 This embodiment conducts comparative experiments from multiple dimensions, including illumination adaptability, obstacle recognition rate, relocation, and power consumption, to verify the performance of the AMR precise positioning and obstacle recognition method in Embodiment 1.

[0116] I. Light Adaptability Experiment Experimental conditions: Test site 500 m 2 Simulated warehouse environment (including shelves, columns, and aisles); prototype of the invention (16-line LiDAR + 2 × single-line LiDAR); comparison scheme is vision-laser fusion scheme (same level LiDAR + binocular camera + infrared supplementary light); lighting conditions cover 0 lux (complete darkness), 5 lux, 50 lux, and 200 lux.

[0117] Experimental results: The positioning accuracy of this invention is completely consistent (fluctuation <5%) across the entire illumination range (0~200 lux), with a local accuracy of ±3.2 mm and a global accuracy of ±5.0 mm. In contrast, the performance of the vision-laser fusion scheme begins to decline significantly when the illumination is below 50 lux, with the local accuracy dropping to ±35.6 mm at 0 lux.

[0118] II. Obstacle Recognition Rate Experiment Experimental conditions: Test objects included a black cardboard box (reflectivity approximately 8%), dark pedestrian clothing (reflectivity approximately 15%), a gray plastic tray (reflectivity approximately 25%), and a white cardboard box (reflectivity approximately 80%); test distances were 5 m, 10 m, and 15 m; illumination conditions were 0 lux.

[0119] Experimental results: This invention achieved a 98% recognition rate for white cardboard boxes, 95% for gray pallets, 92% for dark-colored pedestrians, and 90% for black cardboard boxes. Traditional pure laser SLAM solutions yielded 92%, 78%, 62%, and 45% respectively. This invention improves the recognition rate for low-reflectivity objects (reflectivity ≤15%) from 45%~62% to 90%~92%, an improvement of over 40%.

[0120] III. Relocation Experiment Experimental conditions: The test scenario is a known map area (500 m). 2 Warehouse); initial pose random offset (displacement 0.5~5 m, rotation 0°~90°); illumination condition 0 lux.

[0121] Experimental results: The repositioning success rate of this invention is 100% (50 / 50), with an average time of 67 ms and a repositioning accuracy of ±8.5 cm. The vision-laser fusion scheme has a success rate of only 12% when supplementary lighting is off and only 64% when supplementary lighting is on.

[0122] IV. Power Consumption Comparison Experiment Experimental conditions: Continuous operation for 2 hours, performing the same navigation task.

[0123] Experimental results: The average power consumption of this invention is 42.3 W (relative value 1.00×); the visual-laser fusion scheme consumes 61.8 W (1.46×) at night (with supplementary lighting on) and 48.5 W (1.15×) during the day (with supplementary lighting off). The nighttime power consumption of this invention is approximately 30% lower than that of the visual fusion scheme in nighttime mode.

[0124] Therefore, this invention fundamentally solves the problem of AMR positioning and obstacle avoidance in nighttime and low-light environments by constructing a completely vision-free pure laser perception architecture and matching it with multi-level rigid geometric feature extraction, temporal differential dynamic filtering, scene adaptive matching, and layered obstacle recognition algorithms.

[0125] The system maintains a consistent positioning accuracy of ±3 mm (local) / ±5 mm (global) across the entire illumination range (0~200 lux measured and verified). It achieves a recognition rate of over 90% for low reflectivity objects (reflectivity ≤15%), and a 100% relocation success rate in dark environments (average time 67 ms). At night, its power consumption is reduced by approximately 30% compared to visual fusion solutions.

[0126] Example 3 This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the AMR precise positioning and obstacle recognition method of the visionless lidar in Embodiment 1.

[0127] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the AMR precise positioning and obstacle recognition method of vision-free LiDAR in Embodiment 1.

[0128] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0130] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for precise positioning and obstacle recognition using AMR without visual LiDAR, characterized in that, The method includes: Extract point cloud data of the target scene, extract multi-level rigid geometric features that are inherent and structurally constant in the target scene from the point cloud data, and construct a lightweight pure topology map based on the multi-level rigid geometric features. Obtain the point cloud of the current frame, remove dynamic points in the point cloud of the current frame based on the temporal difference dynamic filtering rules, and obtain the static observation point cloud; Geometric features are extracted from static observation point clouds and matched with a lightweight pure topological map to obtain an initial pose estimate; Based on the feature sparsity of the static observation points, the target scene is segmented according to the feature sparsity to obtain the scene segmentation result; Based on the scene segmentation results, scene adaptive multi-scale matching is performed to iteratively optimize the initial pose estimation and obtain the optimal global pose. Perform hierarchical obstacle recognition on the current frame point cloud, determine the type of obstacle, and determine the spatial position of each obstacle based on the optimal global position; match obstacle avoidance strategies based on the type and spatial position of the obstacle.

2. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 1, characterized in that, The time-series differential dynamic filtering rule is as follows: Based on a preset sliding window, extract historical point clouds from multiple frames that are adjacent to and continuous with the point cloud of the current frame. By unifying multiple historical point clouds into the same coordinate system as the current frame point cloud, the point cloud temporal sequence is obtained. Calculate the point cloud difference between each two adjacent frames of point cloud in the point cloud time sequence, and determine whether the point cloud difference is within the preset difference threshold. If yes, mark the corresponding point cloud as an overlapping point and remove the overlapping point. If no, mark the corresponding point cloud as a non-overlapping point. Track the spatial coordinate changes of each non-overlapping point within a preset sliding window to obtain the displacement change. When the displacement change continues to exceed the preset fluctuation threshold within the sliding window, the non-overlapping point is marked as a dynamic point. Remove all dynamic points from the current frame point cloud to obtain a static observation point cloud.

3. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 1, characterized in that, Extracting inherent and structurally constant multi-level rigid geometric features from point cloud data in the target scene, including: Calculate the local curvature of each point in the point cloud data, and perform classification on each point based on the local curvature to obtain the first level of global structured rigid features; Based on the multi-degree voxel filtering algorithm, the density gradient of point cloud data is calculated in a local region, the location of curvature abrupt change is identified, and the local micro-geometric abrupt change features of the second level are obtained. Extract the height jump characteristics of point cloud data, estimate the ground height based on the height jump characteristics, and perform grid-by-grid region ground fitting based on the ground height to obtain the fitted ground; The boundary information of the fitted ground is extracted to obtain the third-level ground boundary features; the first-level global structured rigid features, the second-level local micro-geometric mutation features, and the third-level ground boundary features are used as the inherent and structurally constant multi-level rigid geometric features in the target scene.

4. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 3, characterized in that, The nodes of the lightweight pure topology map are used to store node location coordinates, node type, multi-level rigid geometric feature descriptors, and node ID and timestamp; wherein, the node type includes corridor nodes, intersection nodes, work nodes, and landmark nodes. The edges of the lightweight pure topology map are used to store the relative pose transformations between nodes, edge types, and information matrices. The information in the lightweight pure topology map includes the map version number, construction time, coordinate system definition, and feature sparsity distribution.

5. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 1, characterized in that, Based on the feature sparsity of static observation points, the target scene is segmented according to the feature sparsity, including: Count the number of valid feature points and the total number of points in the static observation point cloud; Feature sparsity is generated based on the ratio of effective feature points to total number of points; When the feature sparsity is within a preset threshold range, the target scene is divided into a regular region; When the feature sparsity is greater than the maximum value of the threshold region, the target scene is divided into a task docking area; When the feature sparsity is less than the minimum value of the threshold region, the target scene is divided into an open passage area.

6. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 5, characterized in that, Based on the scene segmentation results, perform scene adaptive multi-scale matching, including: When the scene segmentation result is an open passage area, the initial pose estimation is iteratively optimized based on a coarse-scale global topology matching strategy to obtain the optimal global pose. When the scene segmentation result is the work docking area, the initial pose estimation is iteratively optimized based on the fine-scale local micro-curvature matching strategy to obtain the optimal global pose. When the scene segmentation result is a regular region, the initial pose estimation is iteratively optimized by a fusion strategy based on coarse-scale global topology matching and fine-scale local micro-curvature matching to obtain the optimal global pose.

7. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 6, characterized in that, The coarse-scale global topology matching strategy is as follows: Downsampling is performed on the static observation point cloud to generate a sparse feature point cloud; A target grid probability model is constructed based on a lightweight pure topological map, wherein each grid of the target grid probability model pre-stores the mean vector and covariance matrix of the point cloud; The sparse feature point cloud is projected onto the map coordinate system according to the current transformation parameters to obtain several projection points. The initial value of the current transformation parameters is the initial pose estimate. Based on the target grid probability model, calculate the sum of the probability densities of each projection point in each grid. The current transformation parameters are updated using the Jacobian matrix to maximize the sum of probability densities; Extract the poses of two consecutive iterations. Stop iterating when the change in pose between two iterations reaches a preset threshold to obtain the final pose. Use the final pose as the optimal global pose.

8. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 7, characterized in that, The fine-scale local micro-curvature matching strategy is as follows: Extract acute-angle feature points and planar feature points from the static observation point cloud as the source point cloud, and perform downsampling on the source point cloud to generate a sampled source point cloud; Extract subacute angle feature points and subplane feature points near the current pose from a lightweight pure topology map to generate a target point cloud, and perform downsampling on the target point cloud to generate a sampled target point cloud; where the initial value of the current pose is the initial pose estimate; Based on the KD-tree nearest neighbor search algorithm, the nearest neighbor point is obtained by searching the target point cloud and the nearest neighbor point of each point in the source point cloud. An error function is constructed based on the nearest neighbor point, the sampled source point cloud, and the rotation matrix and translation vector to be solved. The Levenberg-Marquardt algorithm is used to solve for the rotation matrix and translation vector to obtain the optimal rotation matrix and translation vector. The optimal global pose is then determined based on the optimal rotation matrix and translation vector.

9. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 8, characterized in that, The fusion strategy for coarse-scale global topology matching and fine-scale local micro-curvature matching is as follows: A coarse-scale global topology matching strategy is executed at low frequency to obtain the first pose estimate and the corresponding confidence score. The second pose estimate is obtained by performing a fine-scale local micro-curvature matching strategy at high frequency. When the confidence score is greater than the confidence threshold, the first pose estimate is used as the initial value of the current pose of the fine-scale local micro-curvature matching strategy, and the second pose estimate is used as the optimal global pose. When the confidence score is greater than the confidence threshold, fusion weights are generated based on a nonlinear least squares function; The first pose estimate and the second pose estimate are weighted and fused based on the fusion weights to obtain the optimal global pose.

10. The AMR precise positioning and obstacle recognition method using a vision-free lidar according to claim 8, characterized in that, The AMR's lidar includes a main lidar and two auxiliary lidars. The main lidar is a three-dimensional multi-threaded lidar, and the auxiliary lidars are two-dimensional single-threaded lidars. The main lidar is used to acquire three-dimensional point clouds around the AMR, and the two auxiliary lidars are used to acquire two-dimensional point clouds positioned above and below the front of the AMR, respectively. The current frame point cloud is a three-dimensional point cloud around the AMR. Perform hierarchical obstacle recognition on the current frame point cloud, including: Ground separation is performed on the two-dimensional point cloud located in front of the AMR based on the CSF algorithm to obtain the ground point cloud; Negative obstacle detection is performed on the ground point cloud to obtain the ground negative obstacle recognition results; High-altitude suspended object detection is performed based on the two-dimensional point cloud positioned in front of the AMR to obtain high-altitude obstacle recognition results; Based on the ground point cloud, remove the ground points from the current frame point cloud to obtain the current frame point cloud after removal; Density clustering is performed on the point cloud of the current frame after the removal of obstacles to obtain multiple independent point cloud clusters; each point cloud cluster represents a potential obstacle. Geometric parameters are calculated for each point cloud cluster to generate three-dimensional geometric indices for the point cloud clusters. Based on the three-dimensional geometric indices, ground positive obstacle identification results are determined. These ground positive obstacle identification results are used to update the lightweight pure topology map. The number of point clouds in a single frame within a point cloud cluster is counted. If the number of point clouds in a single frame is less than a preset number, the point cloud cluster is marked as a suspected low-reflectivity object. Based on the multi-frame temporal accumulation rule, density accumulation is performed on the point cloud clusters suspected of being low reflectivity objects. Then, the point cloud clusters after density accumulation are subjected to point cloud cluster reflection intensity and spatial density analysis and stability verification to obtain the low reflectivity object recognition results.

Images