Intelligent navigation method suitable for goods shelf channel and logistics area environment

Abstract

The invention relates to the technical field of logistics robot navigation, in particular to an intelligent navigation method suitable for a goods shelf channel and a logistics area environment. The method comprises the following steps: acquiring point cloud data acquired by AGV vehicle operation, and after noise suppression of the point cloud data, constructing an accumulator based on Manhattan orthogonal constraint of a plant to carry out orthogonal accumulation weighted statistics on a search neighborhood of the point cloud data so as to determine source landmark point cloud features of an AGV vehicle extraction shelf channel; real-time sensing data of a multi-modal sensor is obtained, a real-time multi-sensor fusion model is constructed based on a hierarchical factor graph and belief propagation calculation, correction registration optimization is performed on source landmark point cloud features by using the real-time multi-sensor fusion model and the optical flow distortion amplitude of dynamic constraint, and positioning navigation is provided for the AGV; according to the invention, powerful environment sensing, positioning and navigation capabilities can be provided for the AGV vehicle, so that efficient warehousing operation and logistics task execution are realized.

Description

Technical Field

[0001] The present invention relates to the field of logistics robot navigation technology, and in particular to an intelligent navigation method suitable for shelf channels and logistics area environments. Background Art

[0002] With the rapid development of autonomous driving, robot navigation, environmental perception and other fields, multimodal sensor fusion technology plays an increasingly important role in intelligent systems. Especially in complex dynamic scenes, traditional single sensors often cannot meet the requirements for high precision, real-time and robustness. Multimodal sensors (such as lidar, camera, radar, etc.) can provide the system with richer and more comprehensive environmental information. However, in practical applications, due to the asynchrony, data delay, noise and other problems between sensors, the data fusion of these sensors has become a huge challenge. Especially in dynamic environments, the interference of motion distortion and sensor noise will reduce the accuracy and effectiveness of sensor data, affecting the final tasks such as target recognition and path planning.

[0003] Point cloud data is the main data form output by sensors such as lidar, and its accuracy and quality are directly related to the perception ability of the system. Point cloud registration technology, as a core technology in point cloud data processing, is used to synthesize point cloud data acquired from multiple different perspectives into a unified global coordinate system. In complex dynamic environments, point cloud registration faces influences such as motion distortion, sensor noise, and data delay. Therefore, it is necessary to develop an intelligent navigation method that can achieve efficient and robust point cloud registration, especially to combat noise interference, to solve the above problems. Summary of the invention

[0004] The present invention overcomes the deficiencies of the prior art and provides an intelligent navigation method suitable for shelf aisles and logistics area environments.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A first aspect of the present invention provides an intelligent navigation method applicable to shelf aisles and logistics area environments, comprising the following steps: S102: Obtain point cloud data collected by the laser radar SLAM operation, identify the Q-table degradation evaluation table through the statistical characteristic degradation of the point cloud data, obtain the maximum degradation evaluation value, and perform point cloud noise suppression on the laser radar SLAM based on the in-degree interference degree calculation of the maximum degradation evaluation value; S104: After noise suppression of the point cloud data, an accumulator is constructed based on the Manhattan orthogonal constraint of the factory to perform orthogonal accumulation weighted statistics on the search neighborhood of the point cloud data to determine the source landmark point cloud features of the shelf channel extracted by the LiDAR SLAM; S106: Acquire real-time sensing data from multimodal sensors, build a real-time multi-sensor fusion model based on hierarchical factor graphs and belief propagation calculations, and use the real-time multi-sensor fusion model and the optical flow distortion amplitude of dynamic constraints to correct and optimize the source landmark point cloud features, providing positioning and navigation for AGV; S108: Continuously incrementally learn and update the tree structure of the initial incremental regression tree of the point cloud data according to the operation requirements of the AGV vehicle, and output an offline mapping model of the factory environment.

[0006] More specifically, the step S102 includes the following steps: Obtain the point cloud data and operation log collected during the operation of the LiDAR SLAM, and extract the historical odometer and prior map information of the AGV vehicle through the operation log; According to the historical odometer and prior map information, a Q-table degradation evaluation table corresponding to the original point cloud action and the original point cloud state of the logistics area is constructed, and the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environmental change pattern of the point cloud data are extracted; Based on the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environment change mode, the current point cloud action is selected and executed to generate a new point cloud action, and the degradation identification state of the Q-table degradation evaluation table is continuously updated and iterated through the new point cloud action to obtain the maximum degradation evaluation value; The dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics area environment are obtained, the dynamic object point cloud and the noise point cloud are identified to construct a degraded directed acyclic graph of the point cloud data, and the interference degree of dynamic objects and environmental noise on the point cloud feature extraction is determined based on the in-degree analysis of the maximum degradation assessment value, and the noise of the abnormal point cloud voxel density is suppressed for the lidar SLAM according to the interference degree.

[0007] More specifically, the acquisition of dynamic object point clouds and noise point clouds of AGV vehicles running in the logistics area environment, identification of the dynamic object point clouds and noise point clouds to construct a degraded directed acyclic graph of point cloud data, determination of the interference degree of dynamic objects and environmental noise on point cloud feature extraction based on the maximum degradation assessment value in-degree analysis of the degraded directed acyclic graph, and noise suppression of abnormal point cloud voxel density for the laser radar SLAM according to the interference degree specifically includes the following steps: Obtain the dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics scene, identify the dynamic object point cloud and the noise point cloud to obtain the dynamic object motion mode and environmental noise characteristics, and construct a degenerate directed acyclic graph of the point cloud data based on the dynamic object motion mode and environmental noise characteristics; Construct a statistical empty stack, calculate the in-degree of each point cloud data node in the degraded directed acyclic graph based on the maximum degradation evaluation value, traverse the in-degree of the adjacent nodes of each point cloud data node in the degraded directed acyclic graph and reduce it by 1. If the in-degree of an adjacent node changes to 0, add the node to the statistical empty stack, and finally output the in-degree statistics. According to the in-degree statistics, determine the interference degree of dynamic objects and environmental noise on point cloud feature extraction; Acquire the point cloud space and space specifications of the point cloud data, divide the point cloud space into N point cloud voxel grids based on the space specification parameters, preset the abnormal density based on the interference degree, introduce the grid density algorithm to calculate the density of each point cloud voxel in each point cloud voxel grid, and obtain a number of point cloud voxel densities; If the point cloud voxel density does not exceed the abnormal density, the point cloud voxel corresponding to the point cloud voxel density is extracted and defined as the representative voxel. All representative voxels are integrated to generate filtered representative voxels. Noise suppression is performed on the point cloud data obtained by the lidar SLAM based on the filtered representative voxels.

[0008] More specifically, the step S104 includes the following steps: After noise suppression of the point cloud data, the size, material and glossiness of the shelf and the Manhattan orthogonal constraints of the factory are obtained, and the degradation of the effective point cloud and the noise point cloud is identified based on the maximum degradation evaluation value of the size, material and glossiness to determine the point cloud generation mode of the shelf channel under the point cloud data obtained by laser radar detection; A search neighborhood of point cloud data is constructed based on the normal direction of the point cloud of the shelf channel in the point cloud data. The Manhattan neighborhood distance between each point cloud in the search neighborhood and the neighboring point cloud is calculated according to the point cloud generation mode, and an accumulator with an orthogonal accumulation benchmark is established according to the Manhattan orthogonal constraint. If the Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance, the point cloud is added to the accumulator for weighted calculation of Manhattan orthogonality, and finally the accumulator's most voting point cloud is output. The source landmark point cloud features of the shelf channel extracted by the lidar SLAM are determined based on the most voting point cloud.

[0009] More specifically, the step S106 includes the following steps: Acquire real-time sensing data of a multimodal sensor, and map underlying state variables of the real-time sensing data with respect to complementarity and redundancy to a local layer, a regional layer, and a global layer to construct a hierarchical factor graph; Continuously propagate confidence information from the bottom leaf nodes of the hierarchical factor graph to the upper local layer and propagate the optimized confidence information from the global layer to the lower layer, repeat the steps of local propagation and global propagation until the hierarchical factor graph is in an open-loop state, and obtain a real-time multi-sensor fusion model; A real-time multi-sensor fusion model is used to obtain the target landmark point cloud features of the shelf aisle, and the optical flow constraint equation of the vehicle dynamics model is constructed based on the point cloud data calculated based on the dynamic parameters of the AGV vehicle. The optical flow constraint equation is solved by smooth constraint iteration to update the optical flow and generate the optical flow distortion amplitude. Based on the optical flow distortion amplitude, the target landmark point cloud features and the source landmark point cloud features are corrected, registered and optimized to perform positioning and navigation for the AGV vehicle running in the shelf aisle and logistics area environment.

[0010] More specifically, the real-time multi-sensor fusion model is used to obtain the target landmark point cloud features of the shelf channel, the optical flow constraint equation of the vehicle dynamics model is constructed based on the point cloud data of the dynamic parameters of the AGV vehicle, the optical flow constraint equation is solved by smooth constraint iteration to update the optical flow, and the optical flow distortion amplitude is generated. Based on the optical flow distortion amplitude, the target landmark point cloud features and the source landmark point cloud features are corrected, registered and optimized, and the AGV vehicle is positioned and navigated in the shelf channel and logistics area environment, which specifically includes the following steps: The dynamic parameters of the AGV vehicle are extracted through the historical odometer, and the vehicle dynamics model of the AGV vehicle is obtained based on big data. The vector gradients of adjacent point clouds of the point cloud data at different time sequences are calculated using the dynamic parameters, and the optical flow algorithm is introduced to construct the optical flow constraint equation of the vehicle dynamics model; Iteratively solving the smooth constraint of the vector gradient through the optical flow constraint equation to obtain an updated optical flow, and calculating the optical flow distortion amplitude between the updated optical flow and the initial optical flow; Use the real-time multi-sensor fusion model to obtain the target landmark point cloud features in the shelf channel in a multi-modal sensing environment, find the point closest to the target point cloud features and each source landmark point cloud feature in the point cloud data, and calculate the rigid transformation matrix between the source landmark point cloud features and the target landmark point cloud features; The rigid transformation matrix is ​​optimized based on the optical flow distortion amplitude correction to obtain the corrected rigid transformation matrix. The position of the source landmark point cloud feature is updated through the corrected rigid transformation matrix and the nearest point registration is re-performed to obtain the landmark point cloud feature registration optimization result of the shelf channel. According to the landmark point cloud feature registration optimization result, the AGV vehicle is positioned and navigated in the shelf channel and logistics area environment.

[0011] More specifically, the step S108 includes the following steps: Obtain the operation requirements of the AGV vehicle, build the root node, split the point cloud data into nodes with different root branches starting from the root node, generate the initial incremental regression tree of the sparse point cloud distribution and the law of environmental changes, and preset the minimum average error according to the operation requirements; Obtain new point cloud data that meets the operation requirements of the AGV vehicle, calculate the prediction error between the new point cloud data and the point cloud data of the initial incremental regression tree, re-split the prediction error based on minimizing the average error, update the tree structure of the initial incremental regression tree incrementally, and output the offline mapping model of the factory environment.

[0012] The present invention solves the technical defects existing in the background technology, and the beneficial technical effects of the present invention are: The point cloud data collected by the LiDAR SLAM operation is obtained, and the Q-table degradation evaluation table is identified through the statistical characteristics of the point cloud data to obtain the maximum degradation evaluation value. The point cloud noise of the LiDAR SLAM is suppressed based on the in-degree interference degree calculation of the maximum degradation evaluation value. After the noise suppression of the point cloud data, an accumulator is constructed based on the Manhattan orthogonal constraint of the factory to perform orthogonal accumulation weighted statistics on the search neighborhood of the point cloud data to determine the source landmark point cloud features of the shelf channel extracted by the LiDAR SLAM. The real-time sensor data of the multimodal sensor is obtained, and a real-time multi-sensor fusion model is constructed based on the hierarchical factor graph and the belief propagation calculation. The source landmark point cloud features are corrected and registered and optimized using the real-time multi-sensor fusion model and the optical flow distortion amplitude of the dynamic constraint to provide positioning and navigation for the AGV vehicle. The tree structure of the initial incremental regression tree of the point cloud data is continuously and incrementally learned and updated according to the operation requirements of the AGV vehicle, and an offline mapping model of the factory environment is output. The present invention can accurately extract and align point clouds for AGV vehicles in environments with repeated shelves and open logistics areas, with features duplication and degradation of optical detection sensors such as lidar SLAM, as well as prior map mismatch and random noise in high-dynamic scenarios, providing AGV vehicles with powerful environmental perception, positioning and navigation capabilities, thereby achieving efficient warehousing operations and logistics task execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, drawings of other embodiments can be obtained based on these drawings without paying creative work.

[0014] Figure 1 A first method flow chart of an intelligent navigation method applicable to shelf aisles and logistics area environments is shown; Figure 2 A second method flow chart of an intelligent navigation method suitable for shelf aisles and logistics area environments is shown. DETAILED DESCRIPTION

[0015] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0016] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0017] The first aspect of the present invention provides an intelligent navigation method suitable for shelf aisles and logistics area environments, such as Figure 1 As shown, the following steps are included: S102: Obtain point cloud data collected by the laser radar SLAM operation, identify the Q-table degradation evaluation table through the statistical characteristic degradation of the point cloud data, obtain the maximum degradation evaluation value, and perform point cloud noise suppression on the laser radar SLAM based on the in-degree interference degree calculation of the maximum degradation evaluation value; S104: After noise suppression of the point cloud data, an accumulator is constructed based on the Manhattan orthogonal constraint of the factory to perform orthogonal accumulation weighted statistics on the search neighborhood of the point cloud data to determine the source landmark point cloud features of the shelf channel extracted by the LiDAR SLAM; S106: Acquire real-time sensor data from multimodal sensors, build a real-time multi-sensor fusion model based on hierarchical factor graphs and belief propagation calculations, and use the real-time multi-sensor fusion model and the optical flow distortion amplitude of dynamic constraints to correct and optimize the source landmark point cloud features, providing positioning and navigation for AGV vehicles; S108: Continuously incrementally learn and update the tree structure of the initial incremental regression tree of the point cloud data according to the operation requirements of the AGV vehicle, and output an offline mapping model of the factory environment.

[0018] More specifically, the step S102 includes the following steps: Obtain the point cloud data and operation log collected during the operation of the LiDAR SLAM, and extract the historical odometer and prior map information of the AGV vehicle through the operation log; According to the historical odometer and prior map information, a Q-table degradation evaluation table corresponding to the original point cloud action and the original point cloud state of the logistics area is constructed, and the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environmental change pattern of the point cloud data are extracted; Based on the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environment change mode, the current point cloud action is selected and executed to generate a new point cloud action, and the degradation identification state of the Q-table degradation evaluation table is continuously updated and iterated through the new point cloud action to obtain the maximum degradation evaluation value; The dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics area environment are obtained, the dynamic object point cloud and the noise point cloud are identified to construct a degraded directed acyclic graph of the point cloud data, and the interference degree of dynamic objects and environmental noise on the point cloud feature extraction is determined based on the in-degree analysis of the maximum degradation assessment value, and the noise of the abnormal point cloud voxel density is suppressed for the lidar SLAM according to the interference degree.

[0019] It should be noted that the autonomous navigation of existing AGV vehicles in logistics scenarios relies on LiDAR point cloud data for environmental perception and positioning. However, when passing through open scenes, due to the sparse environmental features, simple geometric structure and reduced recognition of point cloud data, it is easy to cause positioning degradation problems. In addition, when facing dynamic objects and environmental noise, the existing point cloud feature extraction method is difficult to accurately distinguish dynamic objects from static environments because the sensors are greatly affected by environmental factors and the point cloud data is easily interfered by noise. Figure 1 The consistency is reduced and the re-positioning accuracy is reduced, which greatly reduces the stability of feature extraction and affects the navigation stability of the unmanned vehicle. Therefore, it is necessary to evaluate the degradation problem caused by the AGV vehicle passing through the open scene during the operation of the logistics scene, and to suppress the dynamic objects and noise when running in a complex and changeable environment. In this regard, this method constructs a Q-table degradation evaluation table corresponding to the original point cloud action and the original point cloud state of the logistics area based on the historical odometer and prior map information of the AGV vehicle. The Chinese name of the Q-table degradation evaluation table is Q learning degradation evaluation table, which can deal with the state randomness degradation evaluation of point cloud density, distribution characteristics, plane characteristics, corner distribution and environmental change patterns of point cloud data. By learning the long-term cumulative reward corresponding to the Q degradation value that can be obtained by performing a certain action in a certain state, the degradation evaluation and identification of indoor structured scenes can be realized, which can further make the scene navigation degradation quality dialectic based on positioning prior data more accurate and reliable. The Q-table degradation evaluation table contains the status and actions of the historical odometer and prior map information. The current point cloud action is determined by the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environmental change pattern of the point cloud data. Then, the current point cloud action is executed in the Q-table degradation evaluation table to further observe the new status and reward, so that a new point cloud action can be generated to update the Q-table degradation evaluation table. Finally, the degradation identification state with the maximum Q value recorded in the Q-table degradation evaluation table is the best evaluation value of scene degradation. This method can establish a degradation evaluation mechanism for indoor logistics scenes and the suppression effect of adaptive dynamic objects and noise, which helps to identify and quantify the effects of degraded scenes on the ground. Figure 1 The influence of consistency and repositioning accuracy is reduced, the stability and accuracy of feature extraction is improved, and the stability and robustness of AGV vehicles in complex environments are improved.

[0020] More specifically, the acquisition of dynamic object point clouds and noise point clouds of AGV vehicles running in the logistics area environment, identification of the dynamic object point clouds and noise point clouds to construct a degraded directed acyclic graph of point cloud data, determination of the interference degree of dynamic objects and environmental noise on point cloud feature extraction based on the maximum degradation assessment value in-degree analysis of the degraded directed acyclic graph, and noise suppression of abnormal point cloud voxel density for the laser radar SLAM according to the interference degree specifically includes the following steps: Obtain the dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics scene, identify the dynamic object point cloud and the noise point cloud to obtain the dynamic object motion mode and environmental noise characteristics, and construct a degenerate directed acyclic graph of the point cloud data based on the dynamic object motion mode and environmental noise characteristics; Construct a statistical empty stack, calculate the in-degree of each point cloud data node in the degraded directed acyclic graph based on the maximum degradation evaluation value, traverse the in-degree of the adjacent nodes of each point cloud data node in the degraded directed acyclic graph and reduce it by 1. If the in-degree of an adjacent node changes to 0, add the node to the statistical empty stack, and finally output the in-degree statistics. According to the in-degree statistics, determine the interference degree of dynamic objects and environmental noise on point cloud feature extraction; Acquire the point cloud space and space specifications of the point cloud data, divide the point cloud space into N point cloud voxel grids based on the space specification parameters, preset the abnormal density based on the interference degree, introduce the grid density algorithm to calculate the density of each point cloud voxel in each point cloud voxel grid, and obtain a number of point cloud voxel densities; If the point cloud voxel density does not exceed the abnormal density, the point cloud voxel corresponding to the point cloud voxel density is extracted and defined as the representative voxel. All representative voxels are integrated to generate filtered representative voxels. Noise suppression is performed on the point cloud data obtained by the lidar SLAM based on the filtered representative voxels.

[0021] It should be noted that due to the difficulty in feature extraction caused by dynamic objects and environmental noise when unmanned vehicles are running in complex and changeable environments in logistics scenarios, for the suppression of adaptive dynamic objects and noise, this method first identifies the dynamic object point cloud and noise point cloud of AGV vehicles running in logistics scenarios to determine the motion pattern of dynamic objects and the noise characteristics of the environment, thereby further constructing a degraded directed acyclic graph of point cloud data to reveal the degradation trend ranking of point cloud data obtained under dynamic objects and environmental noise; then construct a statistical empty stack to store point cloud data nodes with an in-degree change of 0, and then use the previously obtained maximum degradation evaluation value to calculate the in-degree of each point cloud data node in the degraded directed acyclic graph, thereby evaluating its ranking changes under the influence of dynamic objects and environmental noise in the degradation trend of point cloud data. Subtracting the in-degree of the adjacent node by 1 can reduce the phenomenon of predecessor dependency and the trend interference between adjacent point cloud data. If the in-degree of an adjacent node changes to 0, it means that the order of the point cloud trend of the point cloud data node is seriously changed after being affected by dynamic objects and environmental noise, which can best reflect the degree of interference of dynamic objects and environmental noise on point cloud data, so the node is added to the statistical empty stack. Finally, all the added point cloud data nodes in the statistical empty stack are a cluster that quantifies the degree of interference of dynamic objects and environmental noise on point cloud feature extraction. This method can effectively distinguish the influence of dynamic objects and environmental noise on point cloud feature extraction by using the degraded scene recognition results and prior constraints, so that AGV vehicles can more quickly detect and clarify the noise level of point cloud data extraction in complex and changeable logistics scenarios. Thereby further extracting environmental features stably and accurately.

[0022] It should be noted that after clarifying the degree of interference of dynamic objects and environmental noise on the point cloud features extracted by LiDAR SLAM, it is necessary to suppress the motion mode of the dynamic objects and the noise characteristics of the environment. To this end, this method adopts the method of dividing the point cloud space of the point cloud data into a number of point cloud voxel grids. On the one hand, the specific location of the point cloud anomaly is refined and located, so that the subsequent suppression of dynamic object motion and environmental noise based on the degree of interference is more accurate, and on the other hand, redundant points are reduced and suppressed to improve the operation efficiency. Then, the density of each point cloud voxel in each point cloud voxel grid is calculated. If the point cloud voxel density does not exceed the abnormal density preset based on the degree of interference, it means that the trend sorting of the point cloud after being affected by dynamic objects and environmental noise is normal. Therefore, the point cloud can be used as a regular sampling representative after noise removal to update the original point cloud with noise, so as to obtain the filtered representative voxel, and the filtered representative voxel is used to suppress the noise of the point cloud data obtained by LiDAR SLAM, so that the suppression of dynamic objects and environmental noise is more accurate and stable. This method can adaptively suppress the noise of dynamic objects and the environment, improve the robustness and accuracy of feature extraction, and help enhance the environmental perception capability of unmanned vehicles, thereby improving the safety and reliability of overall operation, enabling unmanned vehicles to perceive and navigate more reliably in complex logistics environments.

[0023] More specifically, the step S104 includes the following steps: After noise suppression of the point cloud data, the size, material and glossiness of the shelf and the Manhattan orthogonal constraints of the factory are obtained, and the degradation of the effective point cloud and the noise point cloud is identified based on the maximum degradation evaluation value of the size, material and glossiness to determine the point cloud generation mode of the shelf channel under the point cloud data obtained by laser radar detection; A search neighborhood of point cloud data is constructed based on the normal direction of the point cloud of the shelf channel in the point cloud data. The Manhattan neighborhood distance between each point cloud in the search neighborhood and the neighboring point cloud is calculated according to the point cloud generation mode, and an accumulator with an orthogonal accumulation benchmark is established according to the Manhattan orthogonal constraint. If the Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance, the point cloud is added to the accumulator for weighted calculation of Manhattan orthogonality, and finally the accumulator's most voting point cloud is output. The source landmark point cloud features of the shelf channel extracted by the lidar SLAM are determined based on the most voting point cloud.

[0024] It should be noted that the existing technology is often disturbed by the repeatability and random noise of shelf features in the relocation task in the shelf scene, resulting in unstable point cloud feature extraction results and affecting the consistency of relocation. Especially in the factory environment, due to the characteristics of shelf size, material and gloss, the laser radar detection point cloud has a non-uniform density distribution, making it difficult to accurately extract key features. In addition, the traditional feature extraction method does not fully consider the Manhattan hypothesis constraints in the scene, resulting in large errors in the feature mapping from point cloud to landmark. In this regard, after suppressing the dynamic objects and environmental noise extracted from the point cloud data, this method performs degradation identification on the degradation evaluation of the scene point cloud data based on the prior information such as the size, material and gloss of the shelf, so as to obtain the point cloud generation mode of the shelf channel under the laser radar detection point cloud data. Considering that there is a Manhattan hypothesis prior inside the factory, that is, the main structural line is orthogonal to the coordinate axis, this method establishes an accumulator with an orthogonal accumulation benchmark according to the Manhattan orthogonal constraint of the factory, which is used for weighted statistics of point, line and surface parameterized point cloud data that meets the Manhattan orthogonal constraint. If the Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance, it means that the point cloud does not have the interference of inconsistent relocation under the repeated shelf scene features and random noise, so the most accurate source landmark point cloud features of the shelf channel can be further determined by the point cloud with the most votes. This method introduces the degraded scene recognition results and combines the orthogonal constraints of the Manhattan hypothesis to more accurately establish the mapping relationship between point clouds and landmark features, improve the adaptability and robustness of point cloud feature extraction, and thus enhance relocation accuracy and consistency.

[0025] It should be noted that the Manhattan orthogonal constraint is a geometric assumption commonly used in fields such as computer vision and robot navigation. It is based on the Manhattan world hypothesis, that is, the main structures in the factory (such as walls, floors, and ceilings) are usually aligned with a set of fixed, mutually orthogonal directions. In robot intelligent navigation three-dimensional reconstruction, pose estimation, and SLAM positioning and mapping tasks, the Manhattan orthogonal constraint can ensure that the detected planes or edges are aligned with the main orthogonal directions, thereby improving the stability and accuracy of the calculation.

[0026] Calculate the Manhattan neighborhood distance between each point cloud in the search neighborhood and the neighboring point cloud. The corresponding distance formula is: ; in, is the Manhattan neighborhood distance between a point cloud and a neighborhood point cloud, is the distance between the point cloud and a certain neighborhood point cloud on the X-axis, is the distance between the point cloud and a neighboring point cloud on the Y axis, It is the distance between a point cloud and a neighboring point cloud on the Z axis.

[0027] According to the Manhattan orthogonal constraint, an accumulator with an orthogonal accumulation benchmark is established. The corresponding construction formula is: ; in, is the cumulative value, is the weighted weight of the point cloud to the neighboring point cloud, It is the area point cloud whose Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance.

[0028] If the Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance, the point cloud is added to the accumulator for weighted calculation of Manhattan orthogonality. The corresponding weight calculation formula is: ; in, is the weighting function, is the weight distribution adjustment factor.

[0029] Through the above formula steps, it is possible to construct an accumulator based on the Manhattan orthogonal constraint of the factory to orthogonally accumulate the search neighborhood effect of the weighted statistical point cloud data, so that the robot can use the Manhattan orthogonal constraint to match the line features in the image with the orthogonal directions in the world coordinate system during intelligent navigation in the factory, so as to reduce the estimation error and improve the spatial structure accuracy and stability of the point cloud feature extraction.

[0030] More specifically, the step S106 includes the following steps: Acquire real-time sensing data of a multimodal sensor, and map underlying state variables of the real-time sensing data with respect to complementarity and redundancy to a local layer, a regional layer, and a global layer to construct a hierarchical factor graph; Continuously propagate confidence information from the bottom leaf nodes of the hierarchical factor graph to the upper local layer and propagate the optimized confidence information from the global layer to the lower layer, repeat the steps of local propagation and global propagation until the hierarchical factor graph is in an open-loop state, and obtain a real-time multi-sensor fusion model; A real-time multi-sensor fusion model is used to obtain the target landmark point cloud features of the shelf aisle, and the optical flow constraint equation of the vehicle dynamics model is constructed based on the point cloud data calculated based on the dynamic parameters of the AGV vehicle. The optical flow constraint equation is solved by smooth constraint iteration to update the optical flow and generate the optical flow distortion amplitude. Based on the optical flow distortion amplitude, the target landmark point cloud features and the source landmark point cloud features are corrected, registered and optimized to perform positioning and navigation for the AGV vehicle running in the shelf aisle and logistics area environment.

[0031] It should be noted that when AGV vehicles register point cloud data, there are problems such as asynchrony and random errors between multimodal sensors in complex dynamic scenes, which leads to inconsistency in the registration of point cloud data in time and space, and is prone to point cloud registration errors; secondly, the complementarity and redundancy of sensor data under different working conditions such as degradation, non-degradation and high-speed motion have not been fully explored, and it is difficult to adapt to complex environmental changes. The existing fusion method does not adequately model noise characteristics and data delays, resulting in reduced system robustness. In addition, the point cloud data collected by the current AGV automatic guided vehicle during driving is easily affected by changes in its own motion state, resulting in point cloud distortion. Point cloud data with different timestamps may be spatially misaligned, affecting the accuracy of environmental perception and map construction. Therefore, it is necessary to construct a real-time multi-sensor fusion model that guarantees the complementarity and redundancy of multi-sensor data in scenes such as degradation, non-degradation and high speed, as well as a correction mechanism for point cloud data distortion under different motion states, so as to examine the noise characteristics and data delays of different sensors in dynamic environments and explore point cloud data correction under dynamic constraints.

[0032] It should be noted that for the construction of a real-time multi-sensor fusion model, this method maps the underlying state variables of the real-time sensor data of multimodal sensors about complementarity and redundancy to different levels to form a hierarchical factor graph with a hierarchical structure, so that the local reasoning and global optimization between the data of different sensors can be carried out separately, reducing asynchronous and random errors. Then, local belief propagation is performed, that is, the confidence information is propagated from the bottom leaf nodes of the hierarchical factor graph to the upper local layer, and local information fusion is performed between the bottom sensors to improve the reliability of local state estimation. Then, the optimized confidence information is propagated from the global layer to the lower layer, so that the local estimation is adjusted under the global constraint, and global consistency optimization is achieved, which improves the overall fusion synchronization consistency of different sensor data and avoids the generation of local random errors during point cloud registration. If the factor graph has a closed-loop structure, it means that the data fusion overhead of multi-sensors is large and it is difficult to guarantee the accuracy of large-scale data fusion. Therefore, the factor graph is in an open-loop state during multiple rounds of iterative optimization, and finally a real-time multi-sensor fusion model based on hierarchical factor graph and belief propagation can be generated. This method can establish a real-time multi-sensor data fusion mechanism based on factor graph and belief propagation model, make full use of the information complementarity between multimodal sensors, improve the accuracy and stability of sensor data fusion in dynamic environments, and eliminate asynchronous and random errors between multimodal sensors in complex dynamic scenes.

[0033] More specifically, the real-time multi-sensor fusion model is used to obtain the target landmark point cloud features of the shelf channel, the optical flow constraint equation of the vehicle dynamics model is constructed based on the point cloud data calculated based on the dynamic parameters of the AGV vehicle, the optical flow constraint equation is solved by smooth constraint iteration to update the optical flow, and the optical flow distortion amplitude is generated. Based on the optical flow distortion amplitude, the target landmark point cloud features and the source landmark point cloud features are corrected and registered and optimized, so as to locate and navigate the AGV vehicle in the shelf channel and logistics area environment. Figure 2 As shown, the specific steps include: S202: extracting dynamic parameters of the AGV vehicle through the historical odometer, obtaining a vehicle dynamics model of the AGV vehicle based on big data, calculating vector gradients of adjacent point clouds of the point cloud data at different time sequences using the dynamic parameters, and introducing an optical flow algorithm to construct an optical flow constraint equation of the vehicle dynamics model; S204: performing iterative solution of the smoothness constraint on the vector gradient through an optical flow constraint equation to obtain an updated optical flow, and calculating an optical flow distortion amplitude between the updated optical flow and the initial optical flow; S206: using a real-time multi-sensor fusion model to obtain target landmark point cloud features in the shelf channel in a multi-modal sensing environment, finding the closest point between the target point cloud features and each source landmark point cloud feature in the point cloud data, and calculating a rigid transformation matrix between the source landmark point cloud features and the target landmark point cloud features; S208: Optimize the rigid transformation matrix based on the optical flow distortion amplitude correction to obtain a corrected rigid transformation matrix, update the position of the source landmark point cloud feature through the corrected rigid transformation matrix and re-perform the nearest point registration to obtain the landmark point cloud feature registration optimization result of the shelf channel, and perform positioning and navigation for the AGV vehicle running in the shelf channel and logistics area environment based on the landmark point cloud feature registration optimization result.

[0034] It should be noted that, in order to solve the problem of point cloud data distortion caused by motion changes during the driving process of AGV, this method extracts the dynamic parameters of the AGV vehicle based on the historical odometer to calculate the vector gradients of adjacent point clouds of different time series of the point cloud data. The vector gradient contains the spatial gradient and time gradient of the point cloud data, which is used to capture the dynamic information of the point cloud data changing over time. Then, the optical flow constraint equation of the vehicle dynamics model is constructed according to the vector gradient, so that the distortion mode of the point cloud data under different motion states can be analyzed. The smooth constraint iterative solution of the optical flow constraint equation is to ensure that the optical flow is consistent with the grayscale change of the point cloud, which helps to capture the real motion trajectory of the AGV vehicle in the image. The optical flow estimation after each iteration optimization becomes more accurate and can express the smoothness of the point cloud data. The distortion amplitude of the point cloud data can be further revealed based on the updated optical flow. Finally, the optical flow distortion amplitude is the specific quantification of the point cloud data distortion, which is also an important basis for correction. This method can be used to explore point cloud correction under dynamic constraints, which helps to improve the spatial consistency of point cloud data, effectively compensate for errors caused by motion, improve the perception accuracy of AGV vehicles in complex environments, and provide reliable data support for high-precision map construction and autonomous navigation. In vehicle dynamics, optical flow is not only affected by the point cloud gradient, but also related to the movement of the AGV vehicle itself. According to rigid body kinematics, optical flow motion can be derived from the vehicle's dynamic model, and the corresponding optical flow expression formula is: ; Among them, u and v are optical flow components, Z is the depth from the coordinate center point to the laser radar, f is the sensing distance of the laser radar, ω is the angular velocity of the AGV vehicle, X and Y are the relative positions of the scene currently observed by the vehicle, V is the translation velocity component of the vehicle in the world coordinate system, and x and y are the directional velocities.

[0035] Substitute the optical flow expression of the vehicle dynamics model into the optical flow algorithm to derive the corresponding optical flow equation: ; in, is the time gradient.

[0036] During the point cloud acquisition process, since the AGV vehicle follows the movement of the sensor under the vehicle dynamics, the point clouds sampled at different times will be spatially inconsistent, resulting in distortion. The core reason why the optical flow algorithm can be used for point cloud distortion correction is its motion estimation capability, especially when using SLAM lidar or RGB-D camera, it can calculate the motion vector of each point in the point cloud of the acquired image, estimate the true position of the spatial point cloud, and thus compensate for the error caused by motion. The optical flow algorithm is not limited to the optical flow estimation of image pixels, it can also be used for global point cloud motion correction of rigid body motion of unmanned vehicles, and can also be used to capture local non-rigid correction of point clouds in complex motions such as acceleration and rotation. Therefore, the optical flow algorithm can estimate the non-uniform distribution of point clouds under vehicle dynamics constraints to calculate point cloud distortion, and can be combined with IMU, SLAM or deep learning models to improve the accuracy of distortion correction, making point cloud distortion correction more consistent and accurate.

[0037] It should be noted that the existing point cloud feature registration technology is often affected by the dynamic changes of the environment and sensor noise in practical applications, resulting in registration errors. These errors mainly come from sensor errors and environmental changes. Especially in complex or dynamic environments, traditional registration methods are difficult to cope with the uncertainties brought by the environment and sensors. Therefore, the existing registration algorithms have poor adaptability and robustness in dynamic environments, and are prone to large misregistration errors, affecting the accuracy and reliability of point cloud data. In this regard, this method uses a real-time multi-sensor fusion model to obtain the target landmark point cloud features of the shelf channel in a multimodal sensing environment, and then finds the point closest to the target point cloud features and each source landmark point cloud feature in the point cloud data, so that the source point cloud is closer to the point in the target point cloud to establish a one-to-one correspondence, thereby minimizing the error of the final point cloud data registration result. Then the rigid transformation matrix between the source landmark point cloud features and the target landmark point cloud features is calculated. Since the point cloud data is distorted before, this method uses the calculated optical flow distortion amplitude to correct the rigid transformation matrix, thereby providing more accurate rotation and translation transformation for registration, especially in non-rigid or noisy point cloud data, and more reliable registration results can be obtained. Finally, the position of the source landmark point cloud features is updated by the corrected rigid transformation matrix and the nearest point registration is re-performed, so that the AGV vehicle can be more accurately positioned and navigated in the shelf channel and logistics area environment. This method can integrate multiple sources of information, effectively reduce the propagation and amplification of errors, and thus improve the accuracy and stability of point cloud registration, so that the optimization strategy of dynamic constraints can better handle the uncertainty caused by environmental changes and sensor noise in practical applications, and improve the robustness of the registration algorithm. Among them, the rigid transformation matrix is ​​a matrix used to describe the position and direction transformation of an object in space. The rigid transformation retains the shape and size of the object and is usually used to represent translation and rotation operations without scaling or deformation.

[0038] Before rigid body transformation, the centroid between the source landmark point cloud features and the target landmark point cloud features needs to be calculated. The corresponding formula is: ; in, is the centroid of the source landmark point cloud, is the centroid of the target landmark point cloud, n is the total number of points in the point cloud, , For point cloud.

[0039] Then the covariance matrix between the source landmark point cloud features and the target landmark point cloud features is given by: ; Where H is the covariance matrix, , is a column vector, is a row vector and T is a rigid transformation matrix.

[0040] Finally, the rigid transformation matrix can be generated according to the rotation and translation of the covariance matrix. The corresponding rigid transformation matrix formula is: ; Where R is the rotation matrix and t is the translation vector. The rigid transformation matrix is ​​a 4×4 matrix that contains the rotation matrix R and the translation vector t.

[0041] More specifically, the step S108 includes the following steps: Obtain the operation requirements of the AGV vehicle, build the root node, split the point cloud data into nodes with different root branches starting from the root node, generate the initial incremental regression tree of the sparse point cloud distribution and the law of environmental changes, and preset the minimum average error according to the operation requirements; Obtain new point cloud data that meets the operation requirements of the AGV vehicle, calculate the prediction error between the new point cloud data and the point cloud data of the initial incremental regression tree, re-split the prediction error based on minimizing the average error, update the tree structure of the initial incremental regression tree incrementally, and output the offline mapping model of the factory environment.

[0042] It should be noted that when dealing with object distribution and occlusion problems in shelf environments, especially in environments where AGV vehicles operate for a long time, the sparsity of point cloud data will affect the accuracy of environmental perception and map construction. It is difficult to effectively deal with dynamic changes and occlusion problems in the environment, resulting in untimely map updates and may affect repositioning accuracy. In addition, traditional point cloud map construction methods lack adaptability to long-term environmental changes and are difficult to meet the high requirements for accuracy and consistency in practical applications of AGV vehicles. Therefore, this method analyzes the distribution characteristics and environmental change laws of sparse point cloud data by splitting point cloud data nodes according to the operation requirements of AGV vehicles, thereby generating an initial incremental regression tree of sparse point cloud distribution and environmental change laws. Each node splitting can divide the point cloud data into more representative subsets, so that each leaf node can more accurately reflect the distribution characteristics and environmental change laws of sparse point cloud data, ensuring the construction accuracy of the offline mapping model. Then, the prediction error between the new point cloud data and the point cloud data of the initial incremental regression tree is calculated based on the new point cloud data that meets the operation requirements. With the arrival of new data, the existing model is updated and adjusted by minimizing the average error increment to meet the operation requirements, rather than retraining the entire tree. This can ensure that the offline mapping of the factory environment meets the operation requirements of the AGV vehicle to the maximum extent, optimize the distribution of objects in the shelf environment of the AGV vehicle and the sparse point cloud phenomenon caused by repeated occlusion, so that the final output of the factory environment offline mapping model is more stable and accurate. This method can effectively improve the consistency and relocation accuracy of the factory offline map, and can also achieve fast and accurate map updates in complex dynamic environments, ensuring that the AGV vehicle has stable and reliable navigation capabilities in long-term operation.

[0043] In addition, the intelligent navigation method applicable to shelf aisles and logistics area environments further includes the following steps: Obtain an offline mapping model of the plant environment when the AGV vehicle is running for a long time at different time scales, continuously accumulate and calculate the deviation of the offline mapping model of the plant environment based on different time scales, obtain the cumulative model deviation, use the cumulative model deviation to update the historical minimum cumulative deviation of the previous time scale, and generate detection statistics; If the detection statistic is greater than the preset detection statistic, the mean is recalculated or the statistic is reset. If it is less than the preset detection statistic, the cumulative calculation is continued to obtain the online long-term relocation result of the sparse point cloud data. The principal component analysis algorithm is introduced to calculate the normal of the point cloud data, and the angle between each point cloud and the adjacent point cloud is calculated to obtain the normal angle value, and the normal angle threshold is preset according to the online long-term relocation result; If the normal angle value is greater than the normal angle threshold, the point cloud normal of the normal angle value is flipped, and the above normal angle flipping steps of adjacent point clouds are repeated until all point clouds generate new normals to obtain a local rejection point cloud group.

[0044] It should be noted that in the application of sparse point cloud maps, especially in the long-term operation process, map mismatch and drift often occur as the scene changes. Due to environmental changes, sensor errors and uncertainty in data collection, it is difficult to effectively respond to these changes, resulting in a gradual decrease in map accuracy and stability, making it difficult to detect and correct these drift problems in real time, affecting the long-term reliability and adaptability of the system. In addition, due to the interference of dynamic objects, the accuracy and stability of point cloud data are affected. This instability leads to mismatching of point cloud features, which affects the subsequent mapping and positioning accuracy. In this regard, this method cumulatively calculates the model deviation of the offline mapping model of the factory environment when the AGV vehicle runs for a long time on different time scales, and identifies whether there are abnormal fluctuations or changes in the long-term repositioning in the point cloud data; and uses the accumulated model deviation to update the historical minimum cumulative deviation of the previous time scale, so that each time the detection statistic is calculated, it is helpful to distinguish normal fluctuations from abnormal fluctuations compared with the historical lowest point; thereby detecting the change pattern of the sparse point cloud data on different time scales. If the detection statistic is greater than the preset detection statistic, it means that a long-term repositioning anomaly has occurred in the point cloud data, so it is necessary to recalculate the mean or reset the statistic. Otherwise, the cumulative calculation continues, so as to eliminate the anomaly of online long-term repositioning. Then, this method obtains the normal of the point cloud data and calculates the normal angle value between each point cloud and the adjacent point cloud. If the normal angle value is greater than the normal angle threshold preset in the online long-term relocation result, it means that the sparse point cloud data of the normal angle value has a mismatch in the dynamic environment. Therefore, the point cloud normal of the normal angle value can be further flipped and the normal angle flipping step of the adjacent point clouds can be repeated until all point clouds generate new normals to obtain the rejection point cloud group of local mismatch in the point cloud data. This method can use the output results of the offline mapping model to introduce an online environmental change monitoring mechanism, explore the long-term stability of point cloud features under time consistency constraints, and significantly improve the stability, accuracy and robustness of sparse point cloud maps in long-term operation, thereby solving the mismatch and drift problems existing in the prior art; at the same time, a local point cloud rejection algorithm is designed to eliminate the effect of mismatch feature input, eliminate unreliable point cloud data input caused by dynamic objects or mismatch features, and effectively improve the stability of point cloud feature extraction and matching.

[0045] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An intelligent navigation method suitable for shelf aisles and logistics area environments, characterized in that: The following steps are involved: S102: Obtain point cloud data collected by the laser radar SLAM operation, identify the Q-table degradation evaluation table through the statistical characteristic degradation of the point cloud data, obtain the maximum degradation evaluation value, and perform point cloud noise suppression on the laser radar SLAM based on the in-degree interference degree calculation of the maximum degradation evaluation value; S104: After noise suppression of the point cloud data, an accumulator is constructed based on the Manhattan orthogonal constraint of the factory to perform orthogonal accumulation weighted statistics on the search neighborhood of the point cloud data to determine the source landmark point cloud features of the shelf channel extracted by the LiDAR SLAM; S106: Acquire real-time sensing data from multimodal sensors, build a real-time multi-sensor fusion model based on hierarchical factor graphs and belief propagation calculations, and use the real-time multi-sensor fusion model and the optical flow distortion amplitude of dynamic constraints to correct and optimize the source landmark point cloud features, providing positioning and navigation for AGV; S108: Continuously incrementally learn and update the tree structure of the initial incremental regression tree of the point cloud data according to the operation requirements of the AGV vehicle, and output an offline mapping model of the factory environment.

2. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 1, characterized in that: The step S102 specifically includes the following steps: Obtain the point cloud data and operation log collected during the operation of the LiDAR SLAM, and extract the historical odometer and prior map information of the AGV vehicle through the operation log; According to the historical odometer and prior map information, a Q-table degradation evaluation table corresponding to the original point cloud action and the original point cloud state of the logistics area is constructed, and the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environmental change pattern of the point cloud data are extracted; Based on the point cloud density, distribution characteristics, plane characteristics, corner point distribution and environment change mode, the current point cloud action is selected and executed to generate a new point cloud action, and the degradation identification state of the Q-table degradation evaluation table is continuously updated and iterated through the new point cloud action to obtain the maximum degradation evaluation value; The dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics area environment are obtained, the dynamic object point cloud and the noise point cloud are identified to construct a degraded directed acyclic graph of the point cloud data, and the interference degree of dynamic objects and environmental noise on the point cloud feature extraction is determined based on the in-degree analysis of the maximum degradation assessment value, and the noise of the abnormal point cloud voxel density is suppressed for the lidar SLAM according to the interference degree.

3. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 2, characterized in that: The method obtains the dynamic object point cloud and the noise point cloud of the AGV vehicle running in the logistics area environment, identifies the dynamic object point cloud and the noise point cloud to construct a degraded directed acyclic graph of the point cloud data, analyzes the degraded directed acyclic graph based on the maximum degradation assessment value in-degree to determine the interference degree of the dynamic object and the environmental noise on the point cloud feature extraction, and performs noise suppression of the abnormal point cloud voxel density on the laser radar SLAM according to the interference degree, specifically including the following steps: Obtain the dynamic object point cloud and noise point cloud of the AGV vehicle running in the logistics scene, identify the dynamic object point cloud and the noise point cloud to obtain the dynamic object motion mode and environmental noise characteristics, and construct a degenerate directed acyclic graph of the point cloud data based on the dynamic object motion mode and environmental noise characteristics; Construct a statistical empty stack, calculate the in-degree of each point cloud data node in the degraded directed acyclic graph based on the maximum degradation evaluation value, traverse the in-degree of the adjacent nodes of each point cloud data node in the degraded directed acyclic graph and reduce it by 1. If the in-degree of an adjacent node changes to 0, add the node to the statistical empty stack, and finally output the in-degree statistics. According to the in-degree statistics, determine the interference degree of dynamic objects and environmental noise on point cloud feature extraction; Acquire the point cloud space and space specifications of the point cloud data, divide the point cloud space into N point cloud voxel grids based on the space specification parameters, preset the abnormal density based on the interference degree, introduce the grid density algorithm to calculate the density of each point cloud voxel in each point cloud voxel grid, and obtain a number of point cloud voxel densities; If the point cloud voxel density does not exceed the abnormal density, the point cloud voxel corresponding to the point cloud voxel density is extracted and defined as the representative voxel. All representative voxels are integrated to generate filtered representative voxels. Noise suppression is performed on the point cloud data obtained by the lidar SLAM based on the filtered representative voxels.

4. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 1, characterized in that: The step S104 specifically includes the following steps: After noise suppression of the point cloud data, the size, material and glossiness of the shelf and the Manhattan orthogonal constraints of the factory are obtained, and the degradation of the effective point cloud and the noise point cloud is identified based on the maximum degradation evaluation value of the size, material and glossiness to determine the point cloud generation mode of the shelf channel under the point cloud data obtained by laser radar detection; A search neighborhood of point cloud data is constructed based on the normal direction of the point cloud of the shelf channel in the point cloud data. The Manhattan neighborhood distance between each point cloud in the search neighborhood and the neighboring point cloud is calculated according to the point cloud generation mode, and an accumulator with an orthogonal accumulation benchmark is established according to the Manhattan orthogonal constraint. If the Manhattan neighborhood distance is less than the preset Manhattan neighborhood distance, the point cloud is added to the accumulator for weighted calculation of Manhattan orthogonality, and finally the accumulator's most voting point cloud is output. The source landmark point cloud features of the shelf channel extracted by the lidar SLAM are determined based on the most voting point cloud.

5. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 1, characterized in that: The step S106 specifically includes the following steps: Acquire real-time sensing data of a multimodal sensor, and map underlying state variables of the real-time sensing data with respect to complementarity and redundancy to a local layer, a regional layer, and a global layer to construct a hierarchical factor graph; Continuously propagate confidence information from the bottom leaf nodes of the hierarchical factor graph to the upper local layer and propagate the optimized confidence information from the global layer to the lower layer, repeat the steps of local propagation and global propagation until the hierarchical factor graph is in an open-loop state, and obtain a real-time multi-sensor fusion model; A real-time multi-sensor fusion model is used to obtain the target landmark point cloud features of the shelf aisle, and the optical flow constraint equation of the vehicle dynamics model is constructed based on the point cloud data calculated based on the dynamic parameters of the AGV vehicle. The optical flow constraint equation is solved by smooth constraint iteration to update the optical flow and generate the optical flow distortion amplitude. Based on the optical flow distortion amplitude, the target landmark point cloud features and the source landmark point cloud features are corrected, registered and optimized to perform positioning and navigation for the AGV vehicle running in the shelf aisle and logistics area environment.

6. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 5, characterized in that: The method uses a real-time multi-sensor fusion model to obtain the target landmark point cloud features of the shelf channel, calculates the point cloud data based on the dynamic parameters of the AGV vehicle to construct the optical flow constraint equation of the vehicle dynamics model, solves the optical flow constraint equation by smooth constraint iteration to update the optical flow, generates the optical flow distortion amplitude, and corrects and aligns the target landmark point cloud features with the source landmark point cloud features based on the optical flow distortion amplitude, so as to locate and navigate the AGV vehicle in the shelf channel and logistics area environment, specifically including the following steps: The dynamic parameters of the AGV vehicle are extracted through the historical odometer, and the vehicle dynamics model of the AGV vehicle is obtained based on big data. The vector gradients of adjacent point clouds of the point cloud data at different time sequences are calculated using the dynamic parameters, and the optical flow algorithm is introduced to construct the optical flow constraint equation of the vehicle dynamics model; Iteratively solving the smooth constraint of the vector gradient through the optical flow constraint equation to obtain an updated optical flow, and calculating the optical flow distortion amplitude between the updated optical flow and the initial optical flow; Use the real-time multi-sensor fusion model to obtain the target landmark point cloud features in the shelf channel in a multi-modal sensing environment, find the point closest to the target point cloud features and each source landmark point cloud feature in the point cloud data, and calculate the rigid transformation matrix between the source landmark point cloud features and the target landmark point cloud features; The rigid transformation matrix is ​​optimized based on the optical flow distortion amplitude correction to obtain the corrected rigid transformation matrix. The position of the source landmark point cloud feature is updated through the corrected rigid transformation matrix and the nearest point registration is re-performed to obtain the landmark point cloud feature registration optimization result of the shelf channel. According to the landmark point cloud feature registration optimization result, the AGV vehicle is positioned and navigated in the shelf channel and logistics area environment.

7. The intelligent navigation method applicable to shelf aisles and logistics area environments according to claim 1, characterized in that: The step S108 specifically includes the following steps: Obtain the operation requirements of the AGV vehicle, build the root node, split the point cloud data into nodes with different root branches starting from the root node, generate the initial incremental regression tree of the sparse point cloud distribution and the law of environmental changes, and preset the minimum average error according to the operation requirements; Obtain new point cloud data that meets the operation requirements of the AGV vehicle, calculate the prediction error between the new point cloud data and the point cloud data of the initial incremental regression tree, re-split the prediction error based on minimizing the average error, update the tree structure of the initial incremental regression tree incrementally, and output the offline mapping model of the factory environment.

Images