Obstacle map marking method and system combined with unmanned vehicle

Abstract

The present invention provides an obstacle map marking method and system in combination with an unmanned vehicle, which relates to the field of map marking technology, including determining the movement direction and movement path of a mobile unmanned vehicle according to the current location information of the mobile unmanned vehicle and preset target location information; determining the location information of static obstacles by a laser radar scanning device installed on the mobile unmanned vehicle; performing dynamic obstacle target detection on the dynamic point cloud data by a pre-constructed target detection model, and tracking the dynamic obstacles according to a target tracking algorithm to obtain the movement information of the dynamic obstacles; planning a local path for avoiding the static obstacles and dynamic obstacles in real time in the movement direction of the mobile unmanned vehicle by an improved artificial potential field method; and generating a motion control instruction for the mobile unmanned vehicle according to the local path.

Description

Technical Field

[0001] The present invention relates to map marking technology, and in particular to an obstacle map marking method and system combined with an unmanned vehicle. Background Art

[0002] With the rapid development of artificial intelligence and autonomous driving technology, unmanned vehicles have become one of the hot research directions. Unmanned vehicles can autonomously navigate in complex road environments, and their safety and reliability have attracted widespread attention. In order to achieve autonomous decision-making and path planning for unmanned vehicles, on-board sensors are required to perceive the surrounding environment in real time, obtain information such as the location and speed of obstacles, and build a dynamic environment map.

[0003] In the existing technology, unmanned vehicles mainly use laser radar, millimeter wave radar, visual sensors and other devices to perceive the surrounding environment and detect static and dynamic obstacles. These sensors can scan the surrounding area at a certain frequency to obtain information such as the distance and angle of obstacles. However, due to the different characteristics and installation locations of different sensors, the raw data they obtain usually contains noise, outliers and inconsistencies, making it difficult to directly use them to build accurate environmental maps. Summary of the invention

[0004] The embodiments of the present invention provide an obstacle map marking method and system combined with an unmanned vehicle, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provides an obstacle map marking method combined with an unmanned vehicle, including:

[0007] Determine the direction and path of movement of the mobile unmanned vehicle based on the current location information of the mobile unmanned vehicle and the preset target location information; scan the surrounding environment of the mobile unmanned vehicle in real time through the laser radar scanning device installed on the mobile unmanned vehicle to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment;

[0008] The point cloud data is clustered by a clustering segmentation algorithm, static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles are extracted from the point cloud data, and the position information of the static obstacles is determined according to the clustered static point cloud data; dynamic obstacle target detection is performed on the dynamic point cloud data by a pre-built target detection model, and the dynamic obstacle is tracked according to the target tracking algorithm to obtain the motion information of the dynamic obstacle;

[0009] Based on the position information of static obstacles and the motion information of dynamic obstacles, a local path for avoiding the static obstacles and dynamic obstacles is planned in real time in the moving direction of the mobile unmanned vehicle through an improved artificial potential field method; according to the local path, motion control instructions for the mobile unmanned vehicle are generated, and the motion control instructions include speed control instructions and steering control instructions for the mobile unmanned vehicle, which control the mobile unmanned vehicle to move along the local path until it reaches a preset target position.

[0010] In an optional embodiment,

[0011] Performing dynamic obstacle target detection on the dynamic point cloud data by using a pre-built target detection model includes:

[0012] Processing the dynamic point cloud data based on a pre-built target detection model, wherein the target detection model extracts multi-scale point cloud features of the dynamic point cloud data through a PointNet++ backbone network;

[0013] After the PointNet++ backbone network, a multi-scale interactive attention module is cascaded. The multi-scale interactive attention module contains multiple parallel attention units, which process point cloud features of different scales respectively. The point cloud features at each scale are mapped to a low-dimensional embedding space, and the correlation between the point cloud features at different positions is calculated to obtain an attention map. The attention maps of different scales are adaptively aggregated through the fusion module of the target detection model to obtain multi-scale interactive features.

[0014] The target detection model includes three parallel feature pyramid detection heads, which are responsible for the detection of targets of the first size, the second size and the third size respectively. Each feature pyramid detection head includes a center point prediction branch and a target size prediction branch. The center point prediction branch generates a center point heat map consistent with the multi-scale interactive feature resolution through convolution and upsampling layers. The target size prediction branch estimates the target three-dimensional bounding box size corresponding to each center point heat map through regression.

[0015] The outputs of the three detection heads are post-processed through a non-maximum suppression module to obtain a multi-scale target detection result, and at least one of the category, position, size and orientation information of the dynamic obstacle in the dynamic point cloud data is output.

[0016] In an optional embodiment,

[0017] Tracking the dynamic obstacle according to the target tracking algorithm to obtain the motion information of the dynamic obstacle includes:

[0018] The multi-scale target detection results are input into a multi-target tracking framework based on Kalman filtering, wherein the multi-target tracking framework adds an acceleration component to the target state space, and sets a state transfer matrix and a process noise covariance matrix in the target state space, wherein the state transfer matrix and the process noise covariance matrix contain prior statistical information of the target position, velocity and acceleration, and are used to characterize the nonlinear motion characteristics of the target;

[0019] A two-stage cascade matching strategy is used to associate the multi-scale target detection result at the current moment with the previously acquired tracking trajectory. The two-stage cascade matching strategy includes a preliminary screening based on the cosine similarity of the appearance features and a fine screening based on the IoU geometric metric. A set of candidate matching pairs is obtained through the preliminary screening, and a fine screening is performed on this basis to obtain the final matching result.

[0020] Based on the final matching results and the target prior motion model, the state parameters of each dynamic obstacle, including position, velocity and acceleration, are estimated through the prediction and update process of the extended Kalman filter, and a continuous and smooth tracking trajectory is generated.

[0021] In an optional embodiment,

[0022] After obtaining the final matching result, the method further includes performing motion compensation on the mobile unmanned vehicle:

[0023] The measurement information of the on-board IMU and wheel speed meter of the mobile unmanned vehicle is fused to estimate the motion state of the vehicle, and the transformation relationship between the vehicle coordinate system and the world coordinate system is recursively calculated according to the motion state of the vehicle. The prediction result and the observation result of the tracker of the mobile unmanned vehicle are transformed by using the transformation relationship, and the prediction result and the observation result are unified in the world coordinate system to realize the motion compensation of the vehicle.

[0024] In an optional embodiment,

[0025] Based on the position information of the static obstacles and the motion information of the dynamic obstacles, a local path for avoiding the static obstacles and the dynamic obstacles is planned in real time in the motion direction of the mobile unmanned vehicle through an improved artificial potential field method, including:

[0026] A static Gaussian potential field function is established for the position information of the static obstacle, and a dynamic Gaussian potential field function is established for the motion information of the dynamic obstacle, wherein the peak positions of the static Gaussian potential field function and the dynamic Gaussian potential field function are both determined by the position of the obstacle, the dynamic peak position of the dynamic Gaussian potential field function is offset according to the speed direction of the dynamic obstacle, and the offset of the dynamic peak position of the dynamic Gaussian potential field function is proportional to the product of the moving speed of the dynamic obstacle and a preset sliding time window;

[0027] The dynamic Gaussian potential field functions of all dynamic obstacles are weightedly superimposed, and combined with the static Gaussian potential field function to obtain a composite Gaussian potential field function that describes the environment in which the mobile unmanned vehicle is located, wherein the weight coefficient of each obstacle in the composite Gaussian potential field function is adaptively adjusted according to its type, size, and distance from the mobile unmanned vehicle;

[0028] Determining the function gradient of the composite Gaussian potential field function according to the current position of the mobile unmanned vehicle, wherein the function gradient includes a positive gradient direction and a negative gradient direction, and using the negative gradient direction as the target movement direction of the mobile unmanned vehicle, and the target movement direction guides the mobile unmanned vehicle to stay away from static obstacles and dynamic obstacles;

[0029] Taking the negative gradient direction as the preferred direction, combining the kinematic constraints of the mobile unmanned vehicle and the obstacle boundary constraints, a parameterized curve fitting algorithm is used to generate a local path.

[0030] In an optional embodiment,

[0031] The composite Gaussian potential field function is as follows:

[0032] ;

[0033] The static Gaussian potential field function is as follows:

[0034] ;

[0035] The dynamic Gaussian potential field function is as follows:

[0036] ;

[0037] in, U(p,t) express t The composite Gaussian potential field function at time , U s (p) represents the static Gaussian potential field function, U d (p, t) express t The dynamic Gaussian potential field function at time , represents the position vector of the mobile unmanned vehicle in the two-dimensional plane, represents the position vector of the static obstacle, represents the position vector of the dynamic obstacle, A s , A d represent the amplitude of the static Gaussian potential field function and the amplitude of the dynamic Gaussian potential field function respectively, , Respectively represent the covariance matrix of the static Gaussian potential field function and the dynamic Gaussian potential field function, w i Indicates i The weight coefficient corresponding to the dynamic Gaussian potential field function is n Indicates the number of dynamic obstacles.

[0038] In the second aspect of the embodiments of the present application,

[0039] Provides an obstacle map marking system combined with an unmanned vehicle, including:

[0040] The first unit is used to determine the moving direction and moving path of the mobile unmanned vehicle according to the current location information of the mobile unmanned vehicle and the preset target location information; through the laser radar scanning device installed on the mobile unmanned vehicle, the surrounding environment of the mobile unmanned vehicle is scanned in real time to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment;

[0041] The second unit is used to cluster the point cloud data by a clustering segmentation algorithm, extract static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles from the point cloud data, and determine the position information of the static obstacles according to the clustered static point cloud data; perform dynamic obstacle target detection on the dynamic point cloud data by a pre-built target detection model, and track the dynamic obstacles according to the target tracking algorithm to obtain the motion information of the dynamic obstacles;

[0042] The third unit is used to plan a local path that avoids the static obstacles and dynamic obstacles in real time in the movement direction of the mobile unmanned vehicle based on the position information of the static obstacles and the motion information of the dynamic obstacles through an improved artificial potential field method; according to the local path, generate motion control instructions for the mobile unmanned vehicle, the motion control instructions including speed control instructions and steering control instructions for the mobile unmanned vehicle, and control the mobile unmanned vehicle to move along the local path until it reaches a preset target position.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] According to a fourth aspect of the embodiments of the present invention,

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] By obtaining the current position and preset target position information of the mobile unmanned vehicle, the direction and path of the mobile unmanned vehicle are determined, providing a basis for subsequent local path planning, which helps the unmanned vehicle to achieve autonomous navigation and decision-making. The laser radar scanning device on the mobile unmanned vehicle is used to scan the surrounding environment in real time to obtain comprehensive and accurate point cloud data, which reflects the location information of static obstacles and dynamic obstacles in the indoor environment, and provides data support for the unmanned vehicle to perceive complex dynamic environments. The point cloud data is processed through the clustering segmentation algorithm to effectively distinguish static obstacles from dynamic obstacles, extract key environmental information, reduce data redundancy and computational complexity, and improve the efficiency and accuracy of environmental perception.

[0051] For dynamic obstacles, the target detection model and target tracking algorithm are used to realize the recognition, positioning and tracking of dynamic obstacles, and obtain the motion information of dynamic obstacles, including position and speed, which provides an important basis for the dynamic obstacle avoidance and decision-making of the unmanned vehicle. Based on the position information of static obstacles and the motion information of dynamic obstacles, the local path is planned in real time through the improved artificial potential field method. The generated path can effectively avoid static and dynamic obstacles, ensure the safety and continuity of the unmanned vehicle's movement, and improve the autonomous navigation ability of the unmanned vehicle in complex dynamic environments.

[0052] According to the planned local path, the speed and steering control instructions of the unmanned vehicle are generated, and the motion control of the unmanned vehicle is realized, so that it can drive smoothly and accurately along the planned path and finally reach the preset target position to complete the autonomous navigation task. The entire technical solution realizes the complete closed loop of the unmanned vehicle from environmental perception, path planning to motion control. The various modules are closely integrated and coordinated with each other, which improves the autonomous navigation performance and safety of the unmanned vehicle in complex dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of an obstacle map marking method for an unmanned vehicle in combination with an embodiment of the present invention;

[0054] Figure 2 The figure is a schematic diagram of the structure of an obstacle map marking system for an unmanned vehicle in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0057] Figure 1 FIG. 1 is a flow chart of an obstacle map marking method for an unmanned vehicle in accordance with an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S101. Determine the direction and path of movement of the mobile unmanned vehicle according to the current location information of the mobile unmanned vehicle and the preset target location information; scan the surrounding environment of the mobile unmanned vehicle in real time through the laser radar scanning device installed on the mobile unmanned vehicle to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment;

[0059] For example, the current position coordinates of the unmanned vehicle in the global coordinate system are obtained in real time by using the GPS positioning module installed on the mobile unmanned vehicle. The coordinates of the target position are read from the preset navigation task, and the target position is usually given by the user or the upper decision module. The current position coordinates and the target position coordinates are passed to the subsequent path planning module as the starting point and end point of the path planning.

[0060] According to the current position and the target position, the movement direction angle of the unmanned vehicle is calculated. With the current position as the starting point and the target position as the end point, an initial movement path is planned in combination with the movement direction angle. The path can be a straight line path, or a smooth curved path can be generated by considering the kinematic constraints of the unmanned vehicle. The planned movement path is used as a reference for subsequent local path planning, and it is adjusted and optimized in real time according to environmental information.

[0061] One or more laser radar scanning devices are installed on the mobile unmanned vehicle, such as single-line laser radar, multi-line laser radar or three-dimensional laser radar. The laser radar emits laser pulses to the surrounding environment at a certain frequency (such as 10Hz) and receives the returned laser echo signal. According to the emission angle, return time and other information of the laser pulse, the three-dimensional coordinates of each laser point are calculated to form a frame of point cloud data. Each point in the point cloud data represents an observation point in the environment and reflects the position information of the object at that point.

[0062] De-noising is performed on the acquired raw point cloud data to remove abnormal points and outliers to improve data quality. Common denoising methods include statistical filtering and radius filtering. Down-sampling is performed on the point cloud data to reduce the amount of data and improve processing efficiency. Common down-sampling methods include voxel grid down-sampling and random down-sampling. The point cloud data is converted to the local coordinate system of the mobile unmanned vehicle, and a local coordinate system is established with the current position of the unmanned vehicle as the origin and the direction of movement as the positive direction of the x-axis. In the local coordinate system, the point cloud data reflects the instantaneous position information of static obstacles and dynamic obstacles in the environment around the unmanned vehicle.

[0063] Using algorithms such as normal vector segmentation and Euclidean distance clustering, the point cloud data is segmented into multiple independent point cloud clusters, each of which represents a potential obstacle. Feature extraction is performed on each point cloud cluster, and its geometric center, bounding box, main direction and other features are calculated for subsequent obstacle identification and tracking. Static obstacles and dynamic obstacles are distinguished based on the characteristics of the point cloud cluster. The position of static obstacles is fixed, while the position of dynamic obstacles changes over time. The point cloud clusters of static obstacles are merged to form a static environment map; the point cloud clusters of dynamic obstacles are processed separately to extract their motion information.

[0064] According to the clustered point cloud data, the location information of static obstacles is determined. For each static obstacle point cloud cluster, its geometric center coordinates are calculated as the location information of the static obstacle. The location information of all static obstacles is combined into a static environment map for subsequent path planning and obstacle avoidance decisions. Static environment maps can be represented by different data structures, such as occupancy grid maps, topological maps, etc. The appropriate map representation method is selected according to the specific application requirements.

[0065] Use pre-trained target detection models (such as SSD, YOLO, etc.) to perform target detection on each dynamic obstacle point cloud cluster and identify different types of dynamic obstacles. According to the target detection results, extract the position, size, orientation and other information of each dynamic obstacle to construct the motion state vector of the dynamic obstacle. Use multi-target tracking algorithms (such as Kalman filtering, particle filtering, etc.) to track and predict the motion state of each dynamic obstacle and estimate its motion trajectory in the future. Use the real-time position, speed, acceleration and other information of the dynamic obstacle as its motion information for subsequent path planning and obstacle avoidance decisions.

[0066] Through the above steps, the mobile unmanned vehicle can obtain the current position information and preset target position information in real time, and determine the initial movement direction and path; at the same time, the point cloud data of the surrounding environment is obtained through the lidar scanning device, and the point cloud data is segmented, clustered and identified, and the position information of static obstacles and the motion information of dynamic obstacles are extracted, providing environmental perception data support for subsequent path planning and obstacle avoidance decisions.

[0067] S102. Clustering the point cloud data by a clustering segmentation algorithm, extracting static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles from the point cloud data, and determining the position information of the static obstacles according to the clustered static point cloud data; performing dynamic obstacle target detection on the dynamic point cloud data by a pre-built target detection model, and tracking the dynamic obstacles according to a target tracking algorithm to obtain motion information of the dynamic obstacles;

[0068] Exemplarily, select a suitable clustering segmentation algorithm, such as DBSCAN, K-means, spectral clustering, etc., and select the algorithm according to the characteristics of the point cloud data and the clustering target. Set the parameters of the clustering algorithm, such as the clustering distance threshold, the clustering point number threshold, etc. These parameters determine the granularity and effect of the clustering. Input the preprocessed point cloud data into the clustering algorithm, and the algorithm will divide the point cloud data into multiple independent point cloud clusters based on the distance or similarity measurement between the points. Each point cloud cluster represents a potential obstacle or object, and the points inside the point cloud cluster have similar features or positions.

[0069] The clustered point cloud clusters are analyzed, and static obstacles and dynamic obstacles are distinguished based on the characteristics of the point cloud clusters. The point cloud clusters of static obstacles remain fixed in the continuous multi-frame point cloud data, while the point cloud clusters of dynamic obstacles change position between different frames. By tracking the position changes of the point cloud clusters in continuous frames, the speed and trajectory of the point cloud clusters are calculated to determine whether they are static obstacles or dynamic obstacles. The point cloud clusters belonging to static obstacles are extracted to form a static point cloud data set; the point cloud clusters belonging to dynamic obstacles are extracted to form a dynamic point cloud data set.

[0070] For each static point cloud cluster, calculate its geometric center coordinates as the location information of the static obstacle. The geometric center coordinates can be obtained by averaging the coordinates of all points in the point cloud cluster; the location information of all static obstacles is combined into a static environment map for subsequent path planning and obstacle avoidance decisions. Static environment maps can be represented by different data structures, such as occupancy grid maps, topological maps, etc., and the appropriate map representation method is selected according to specific application requirements.

[0071] For each dynamic point cloud cluster, its motion parameters such as velocity and acceleration are estimated by tracking its position changes in consecutive frames. Common motion estimation methods include Kalman filtering and particle filtering. By establishing a motion model of dynamic obstacles, their states are recursively estimated and updated. Based on the estimated motion parameters, the motion trajectory and position distribution of dynamic obstacles in the future are predicted to provide a reference for subsequent path planning and obstacle avoidance decisions. The real-time position, velocity, acceleration and other information of dynamic obstacles are used as their motion information and compared with the static environment. Figure 1 Together they constitute complete environmental perception data.

[0072] Through the above steps, the mobile unmanned vehicle can cluster the point cloud data through the clustering segmentation algorithm, extract static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles; determine the location information of static obstacles based on the static point cloud data and build a static environment map; perform motion estimation and prediction on the dynamic point cloud data to obtain the motion information of dynamic obstacles.

[0073] In an optional embodiment,

[0074] Performing dynamic obstacle target detection on the dynamic point cloud data by using a pre-built target detection model includes:

[0075] Processing the dynamic point cloud data based on a pre-built target detection model, wherein the target detection model extracts multi-scale point cloud features of the dynamic point cloud data through a PointNet++ backbone network;

[0076] After the PointNet++ backbone network, a multi-scale interactive attention module is cascaded. The multi-scale interactive attention module contains multiple parallel attention units, which process point cloud features of different scales respectively. The point cloud features at each scale are mapped to a low-dimensional embedding space, and the correlation between the point cloud features at different positions is calculated to obtain an attention map. The attention maps of different scales are adaptively aggregated through the fusion module of the target detection model to obtain multi-scale interactive features.

[0077] The target detection model includes three parallel feature pyramid detection heads, which are responsible for the detection of targets of the first size, the second size and the third size respectively. Each feature pyramid detection head includes a center point prediction branch and a target size prediction branch. The center point prediction branch generates a center point heat map consistent with the multi-scale interactive feature resolution through convolution and upsampling layers. The target size prediction branch estimates the target three-dimensional bounding box size corresponding to each center point heat map through regression.

[0078] The outputs of the three detection heads are post-processed through a non-maximum suppression module to obtain a multi-scale target detection result, and at least one of the category, position, size and orientation information of the dynamic obstacle in the dynamic point cloud data is output.

[0079] For example, PointNet++ is selected as the backbone network of the target detection model to extract multi-scale features of dynamic point cloud data. The PointNet++ network generates point cloud features of multiple scales through layer-by-layer downsampling and feature extraction. The features of each scale represent point cloud representations at different levels of abstraction. The input of the PointNet++ network is dynamic point cloud data, and the output is multi-scale point cloud features, which provide rich feature representations for subsequent target detection tasks.

[0080] After the PointNet++ backbone network, a multi-scale interactive attention module is cascaded to enhance the interaction and fusion between features of different scales. The multi-scale interactive attention module contains multiple parallel attention units, each of which corresponds to a point cloud feature of one scale. For each scale of point cloud features, the attention unit maps them to a low-dimensional embedding space to obtain the feature embedding representation at that scale. The correlation between feature embeddings at different positions is calculated to generate an attention map to capture the spatial dependency of point cloud features. Through the fusion module of the target detection model, attention maps of different scales are adaptively aggregated to obtain multi-scale interactive features, thereby achieving effective fusion of features of different scales.

[0081] The object detection model includes three parallel feature pyramid detection heads, which are responsible for detecting objects of different sizes. The first detection head is responsible for detecting small-sized objects, the second detection head is responsible for detecting medium-sized objects, and the third detection head is responsible for detecting large-sized objects. Each feature pyramid detection head consists of two branches: a center point prediction branch and a target size prediction branch. The center point prediction branch gradually restores the multi-scale interactive features to the same resolution as the input point cloud through convolution and upsampling layers, and generates a corresponding center point heat map. The target size prediction branch estimates the size of the target 3D bounding box corresponding to each center point heat map position through regression, including parameters such as length, width, and height.

[0082] The dynamic point cloud data is input into the constructed object detection model, and the detection results of three scales are obtained after being processed by the PointNet++ backbone network, the multi-scale interactive attention module and the feature pyramid detection head. For the detection results of each scale, a set of candidate object detection boxes are generated through the center point heat map generated by the center point prediction branch and the bounding box size estimated by the target size prediction branch. The candidate detection boxes of the three scales are merged, and the duplicate and redundant detection boxes are removed by the non-maximum suppression (NMS) algorithm to obtain the final multi-scale object detection results. The object detection results include the category (such as pedestrians, vehicles, etc.), position (center coordinates of the three-dimensional bounding box), size (length, width, height of the three-dimensional bounding box) and orientation (direction of the three-dimensional bounding box) of the detected dynamic obstacles.

[0083] Through the above steps, the mobile unmanned vehicle can use the pre-built target detection model to process dynamic point cloud data, extract multi-scale point cloud features through the PointNet++ backbone network, and use the multi-scale interactive attention module to enhance the interaction and fusion of features. Then, three parallel feature pyramid detection heads are used to detect targets of different sizes, and post-process them through non-maximum suppression, and finally output information such as the category, position, size, and orientation of dynamic obstacles, providing important environmental perception data for obstacle avoidance and navigation of the unmanned vehicle.

[0084] In an optional embodiment,

[0085] Tracking the dynamic obstacle according to the target tracking algorithm to obtain the motion information of the dynamic obstacle includes:

[0086] The multi-scale target detection results are input into a multi-target tracking framework based on Kalman filtering, wherein the multi-target tracking framework adds an acceleration component to the target state space, and sets a state transfer matrix and a process noise covariance matrix in the target state space, wherein the state transfer matrix and the process noise covariance matrix contain prior statistical information of the target position, velocity and acceleration, and are used to characterize the nonlinear motion characteristics of the target;

[0087] A two-stage cascade matching strategy is used to associate the multi-scale target detection result at the current moment with the previously acquired tracking trajectory. The two-stage cascade matching strategy includes a preliminary screening based on the cosine similarity of the appearance features and a fine screening based on the IoU geometric metric. A set of candidate matching pairs is obtained through the preliminary screening, and a fine screening is performed on this basis to obtain the final matching result.

[0088] Based on the final matching results and the target prior motion model, the state parameters of each dynamic obstacle, including position, velocity and acceleration, are estimated through the prediction and update process of the extended Kalman filter, and a continuous and smooth tracking trajectory is generated.

[0089] Exemplarily, a multi-target tracking framework based on Kalman filtering is constructed. The target state space is designed, including the position, velocity and acceleration components of the target, to describe the motion state of the dynamic obstacle. The state transfer matrix is ​​defined to characterize the changing relationship of the target state between consecutive time steps, and the dynamic model of the position, velocity and acceleration of the target is considered. The process noise covariance matrix is ​​set to represent the uncertainty and random disturbance in the target motion process, which is used to adjust the prediction and update process of the Kalman filter. The prior statistical information of the target position, velocity and acceleration is introduced into the state transfer matrix and the process noise covariance matrix to better characterize the nonlinear motion characteristics of the target.

[0090] Get the multi-scale target detection results at the current moment, including the target's category, position, size, and orientation. Get the pre-existing tracking tracks from the tracking results at the previous moment, each of which corresponds to a dynamic obstacle. Use a two-stage cascade matching strategy to associate the current detection results with the pre-existing tracking tracks.

[0091] The first stage (initial screening): Calculate the cosine similarity of the appearance features between each detection result and each tracking trajectory, screen out candidate matching pairs with similarity higher than the threshold, and form a set of candidate matching pairs.

[0092] The second stage (fine screening): For each matching pair in the candidate matching pair set, the intersection over union (IoU) between the detection result and the tracking trajectory is calculated, and the matching pair with the largest IoU is selected as the final matching result.

[0093] For each final matched detection result and tracking trajectory pair, the extended Kalman filter is used to estimate the target state. The prediction step of the extended Kalman filter is: based on the target state and state transfer matrix at the previous moment, the target state at the current moment is predicted to obtain the predicted position, velocity and acceleration.

[0094] The update steps of the extended Kalman filter are as follows: the detection result at the current moment is used as the observation value, combined with the predicted target state and the observation noise covariance matrix, the target state is updated through the Kalman gain, and the updated position, velocity and acceleration estimates are obtained.

[0095] For unmatched detection results, it is considered to be a newly appeared target, and a new tracking trajectory is initialized for it, and the initial state estimate is set. For unmatched tracking trajectories, it is considered that the target may be blocked or leave the field of view, and whether to retain or terminate the tracking trajectory is decided according to the strategy. According to the updated target state estimate, the tracking trajectory of each dynamic obstacle is generated to achieve continuous and smooth target motion representation. For each tracked dynamic obstacle, the position, velocity and acceleration information in its state estimate is extracted. The motion information of the dynamic obstacle is encoded and serialized in an appropriate data format (such as JSON, XML, etc.) to facilitate subsequent transmission and processing. The motion information of the dynamic obstacle is sent to other modules of the unmanned vehicle, such as path planning, decision control, etc., for them to make obstacle avoidance and navigation decisions.

[0096] Through the above steps, the mobile unmanned vehicle can use the multi-target tracking framework based on Kalman filtering to track and estimate the state of multi-scale target detection results. By introducing the position, velocity and acceleration information of the target and adding prior statistical information to the state transfer matrix and process noise covariance matrix, the nonlinear motion characteristics of the target can be better characterized. A two-stage cascade matching strategy is used for data association, and a reliable match between the detection results and the tracking trajectory is achieved by combining the appearance feature similarity and geometric measurement. Finally, by extending the prediction and update process of the Kalman filter, the state parameters of each dynamic obstacle are estimated, a continuous and smooth tracking trajectory is generated, and the motion information of the dynamic obstacle is provided to other modules of the unmanned vehicle to support the realization of obstacle avoidance and navigation functions.

[0097] In an optional embodiment,

[0098] After obtaining the final matching result, the method further includes performing motion compensation on the mobile unmanned vehicle:

[0099] The measurement information of the on-board IMU and wheel speed meter of the mobile unmanned vehicle is fused to estimate the motion state of the vehicle, and the transformation relationship between the vehicle coordinate system and the world coordinate system is recursively calculated according to the motion state of the vehicle. The prediction result and the observation result of the tracker of the mobile unmanned vehicle are transformed by using the transformation relationship, and the prediction result and the observation result are unified in the world coordinate system to realize the motion compensation of the vehicle.

[0100] Exemplarily, obtain the measurement information of the vehicle-mounted inertial measurement unit (IMU) on the mobile unmanned vehicle, including angular velocity and acceleration data. Obtain the measurement information of the wheel speed meter on the mobile unmanned vehicle, including wheel speed and vehicle speed data. Synchronize the time and align the coordinate system of the measurement information of the IMU and the wheel speed meter to ensure the consistency of the data in time and space. Use a Kalman filter or other fusion algorithm to fuse the measurement information of the IMU and the wheel speed meter to estimate the motion state of the mobile unmanned vehicle, including parameters such as position, speed, and attitude.

[0101] The vehicle coordinate system is usually defined with the geometric center of the mobile unmanned vehicle as the origin, the vehicle's forward direction as the x-axis, the left side as the y-axis, and the vertical upward as the z-axis. To define the world coordinate system, a fixed global reference system is usually selected, such as the Northeastern Universe (ENU) coordinate system or the global geographic coordinate system (such as WGS84). Based on the estimated motion state of the ego vehicle, the transformation matrix or transformation parameters of the vehicle coordinate system relative to the world coordinate system are recursively calculated, including the translation vector and the rotation matrix. The calculation of the transformation matrix or transformation parameters can use the dead reckoning method to recursively update the transformation relationship between the vehicle coordinate system and the world coordinate system based on the changes in the position, speed, and attitude of the ego vehicle.

[0102] Get the prediction results of the multi-target tracker in the vehicle coordinate system, including the predicted position, velocity, acceleration and other state quantities of each target. Use the transformation relationship between the vehicle coordinate system and the world coordinate system to transform the prediction results of the tracker from the vehicle coordinate system to the world coordinate system. The coordinate transformation can be achieved by multiplying the position vector of the prediction result with the transformation matrix and adding the translation vector. The transformed prediction results represent the predicted state of each target in the world coordinate system, including information such as position, velocity and acceleration.

[0103] Get the multi-scale target detection results at the current moment as the tracker's observation results, including the detection position, size, and orientation of each target. Use the transformation relationship between the vehicle coordinate system and the world coordinate system to transform the tracker's observation results from the vehicle coordinate system to the world coordinate system. Coordinate transformation can be achieved by multiplying the position vector of the observation result with the transformation matrix and adding the translation vector. The transformed observation results represent the detection status of each target in the world coordinate system, including information such as position, size, and orientation.

[0104] The transformed prediction results and observation results are input into the multi-target tracker to perform target tracking and state estimation in the world coordinate system. The multi-target tracker uses a data association algorithm (such as the Hungarian algorithm or the JPDA algorithm) to match the observation results with the existing tracking trajectory based on the prediction results and observation results. For successfully matched observation results and tracking trajectories, a Kalman filter or other estimation algorithm is used to update the target state estimate to obtain information such as the target position, velocity, and acceleration in the world coordinate system. For unmatched observation results, they are considered as newly appeared targets and new tracking trajectories are initialized; for unmatched tracking trajectories, they are considered as disappeared or blocked, and the strategy determines whether to retain or terminate the tracking trajectory.

[0105] Through the above steps, the mobile unmanned vehicle can realize the motion compensation of the vehicle and unify the prediction results and observation results of the tracker in the world coordinate system for processing. First, by fusing the measurement information of the on-board IMU and the wheel speed meter, the motion state of the vehicle is estimated, and the transformation relationship between the vehicle coordinate system and the world coordinate system is recursively calculated. Then, using this transformation relationship, the prediction results and observation results of the tracker are transformed from the vehicle coordinate system to the world coordinate system. Finally, target tracking and state estimation are performed in the world coordinate system to obtain motion information such as the position, velocity and acceleration of dynamic obstacles, providing more accurate and consistent perception information for the decision-making control of the unmanned vehicle.

[0106] S103. Based on the position information of static obstacles and the motion information of dynamic obstacles, a local path for avoiding the static obstacles and dynamic obstacles is planned in real time in the moving direction of the mobile unmanned vehicle through an improved artificial potential field method; according to the local path, motion control instructions for the mobile unmanned vehicle are generated, and the motion control instructions include speed control instructions and steering control instructions for the mobile unmanned vehicle, which control the mobile unmanned vehicle to move along the local path until it reaches a preset target position.

[0107] In an optional embodiment,

[0108] Based on the position information of the static obstacles and the motion information of the dynamic obstacles, a local path for avoiding the static obstacles and the dynamic obstacles is planned in real time in the motion direction of the mobile unmanned vehicle through an improved artificial potential field method, including:

[0109] A static Gaussian potential field function is established for the position information of the static obstacle, and a dynamic Gaussian potential field function is established for the motion information of the dynamic obstacle, wherein the peak positions of the static Gaussian potential field function and the dynamic Gaussian potential field function are both determined by the position of the obstacle, the dynamic peak position of the dynamic Gaussian potential field function is offset according to the speed direction of the dynamic obstacle, and the offset of the dynamic peak position of the dynamic Gaussian potential field function is proportional to the product of the moving speed of the dynamic obstacle and a preset sliding time window;

[0110] The dynamic Gaussian potential field functions of all dynamic obstacles are weightedly superimposed, and combined with the static Gaussian potential field function to obtain a composite Gaussian potential field function that describes the environment in which the mobile unmanned vehicle is located, wherein the weight coefficient of each obstacle in the composite Gaussian potential field function is adaptively adjusted according to its type, size, and distance from the mobile unmanned vehicle;

[0111] Determining the function gradient of the composite Gaussian potential field function according to the current position of the mobile unmanned vehicle, wherein the function gradient includes a positive gradient direction and a negative gradient direction, and using the negative gradient direction as the target movement direction of the mobile unmanned vehicle, and the target movement direction guides the mobile unmanned vehicle to stay away from static obstacles and dynamic obstacles;

[0112] Taking the negative gradient direction as the preferred direction, combining the kinematic constraints of the mobile unmanned vehicle and the obstacle boundary constraints, a parameterized curve fitting algorithm is used to generate a local path.

[0113] Exemplarily, the location information of static obstacles in the environment is obtained, including the coordinates, size, and shape of the obstacles. For each static obstacle, a Gaussian potential field function is constructed, and the peak position of the function coincides with the position of the obstacle. The Gaussian potential field function can be in the form of a two-dimensional or three-dimensional Gaussian distribution, where the peak position represents the center of the obstacle, and the standard deviation or covariance matrix of the function is set according to the size and shape of the obstacle. The static Gaussian potential field function represents the spatial distribution and influence range of static obstacles in the environment. The potential field value at the peak position is the highest, indicating the area that is most unfavorable for the passage of mobile unmanned vehicles.

[0114] Obtain the motion information of dynamic obstacles in the environment, including the position, velocity, and acceleration of the obstacles. For each dynamic obstacle, construct a Gaussian potential field function, and the peak position of the function initially coincides with the position of the obstacle. According to the velocity direction and magnitude of the dynamic obstacle, dynamically offset the peak position of the Gaussian potential field function. The offset calculation formula is: offset = velocity magnitude × preset sliding time window. The form of the dynamic Gaussian potential field function is similar to the static Gaussian potential field function, but the peak position changes dynamically with the movement of the obstacle, indicating the predicted position and impact range of the dynamic obstacle in the future.

[0115] The dynamic Gaussian potential field functions of all dynamic obstacles are weightedly superimposed to obtain a comprehensive dynamic Gaussian potential field function. The static Gaussian potential field function is superimposed with the comprehensive dynamic Gaussian potential field function to obtain a composite Gaussian potential field function that describes the environment in which the mobile unmanned vehicle is located. In the composite Gaussian potential field function, a weight coefficient is assigned to each obstacle, and the size of the weight coefficient is adaptively adjusted according to the type of obstacle (such as pedestrians, vehicles, etc.), size, and distance from the mobile unmanned vehicle. The type and size of the obstacle determine its potential threat to the mobile unmanned vehicle. The closer the obstacle is, the greater the weight is, and the greater the impact on the mobile unmanned vehicle is.

[0116] According to the current position of the mobile unmanned vehicle, the gradient of the composite Gaussian potential field function at that position is calculated. The gradient of the composite Gaussian potential field function includes two directions: positive gradient direction and negative gradient direction. The positive gradient direction points to the direction where the potential field value increases fastest, indicating that it is close to the obstacle; the negative gradient direction points to the direction where the potential field value decreases fastest, indicating that it is far away from the obstacle. The negative gradient direction of the composite Gaussian potential field function is used as the target movement direction of the mobile unmanned vehicle, which guides the mobile unmanned vehicle to stay away from static obstacles and dynamic obstacles and drive towards a safe area with a lower potential field value.

[0117] Taking the negative gradient direction of the composite Gaussian potential field function as the preferred direction, a local path is generated by combining the kinematic constraints of the mobile unmanned vehicle (such as maximum speed, acceleration, steering angle, etc.) and obstacle boundary constraints (such as keeping a safe distance from obstacles). The generation of the local path can use parameterized curve fitting algorithms, such as spline curves and Bezier curves, etc., to optimize the position of the control points so that the generated curve satisfies the kinematic constraints and obstacle boundary constraints and is aligned with the preferred direction as much as possible. The generated local path provides a smooth and safe obstacle avoidance trajectory for the mobile unmanned vehicle, guiding the mobile unmanned vehicle to travel between static obstacles and dynamic obstacles, and adjusting the driving direction and speed in time.

[0118] Through the above steps, the mobile unmanned vehicle can use the improved artificial potential field method to plan a local path to avoid static obstacles and dynamic obstacles in real time. First, a static Gaussian potential field function is established based on the position information of the static obstacle, and a dynamic Gaussian potential field function is established based on the motion information of the dynamic obstacle, and a composite Gaussian potential field function is obtained by weighted superposition. Then, the negative gradient direction of the composite Gaussian potential field function is determined according to the current position of the mobile unmanned vehicle as the target motion direction. Finally, the negative gradient direction is used as the preferred direction, combined with kinematic constraints and obstacle boundary constraints, and a parameterized curve fitting algorithm is used to generate a smooth and safe local path to guide the mobile unmanned vehicle to achieve obstacle avoidance and navigation in complex environments.

[0119] In an optional embodiment,

[0120] The composite Gaussian potential field function is as follows:

[0121] ;

[0122] The static Gaussian potential field function is as follows:

[0123] ;

[0124] The dynamic Gaussian potential field function is as follows:

[0125] ;

[0126] in, U(p,t) express t The composite Gaussian potential field function at time , U s (p) represents the static Gaussian potential field function, U d (p, t) express t The dynamic Gaussian potential field function at time , represents the position vector of the mobile unmanned vehicle in the two-dimensional plane, represents the position vector of the static obstacle, represents the position vector of the dynamic obstacle, A s , A d represent the amplitude of the static Gaussian potential field function and the amplitude of the dynamic Gaussian potential field function respectively, , Respectively represent the covariance matrix of the static Gaussian potential field function and the dynamic Gaussian potential field function, w i Indicates i The weight coefficient corresponding to the dynamic Gaussian potential field function is n Indicates the number of dynamic obstacles.

[0127] Figure 2 FIG. 1 is a schematic diagram of the structure of an obstacle map marking system for an unmanned vehicle according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0128] The first unit is used to determine the moving direction and moving path of the mobile unmanned vehicle according to the current location information of the mobile unmanned vehicle and the preset target location information; through the laser radar scanning device installed on the mobile unmanned vehicle, the surrounding environment of the mobile unmanned vehicle is scanned in real time to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment;

[0129] The second unit is used to cluster the point cloud data by a clustering segmentation algorithm, extract static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles from the point cloud data, and determine the position information of the static obstacles according to the clustered static point cloud data; perform dynamic obstacle target detection on the dynamic point cloud data by a pre-built target detection model, and track the dynamic obstacles according to the target tracking algorithm to obtain the motion information of the dynamic obstacles;

[0130] The third unit is used to plan a local path that avoids the static obstacles and dynamic obstacles in real time in the movement direction of the mobile unmanned vehicle based on the position information of the static obstacles and the motion information of the dynamic obstacles through an improved artificial potential field method; according to the local path, generate motion control instructions for the mobile unmanned vehicle, the motion control instructions including speed control instructions and steering control instructions for the mobile unmanned vehicle, and control the mobile unmanned vehicle to move along the local path until it reaches a preset target position.

[0131] According to a third aspect of the embodiments of the present invention,

[0132] An electronic device is provided, comprising:

[0133] processor;

[0134] a memory for storing processor-executable instructions;

[0135] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0136] According to a fourth aspect of the embodiments of the present invention,

[0137] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0138] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The obstacle map marking method combined with the unmanned vehicle is characterized by: include: Determine the direction and path of movement of the mobile unmanned vehicle based on the current location information of the mobile unmanned vehicle and the preset target location information; scan the surrounding environment of the mobile unmanned vehicle in real time through the laser radar scanning device installed on the mobile unmanned vehicle to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment; The point cloud data is clustered by a clustering segmentation algorithm, static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles are extracted from the point cloud data, and the position information of the static obstacles is determined according to the clustered static point cloud data; dynamic obstacle target detection is performed on the dynamic point cloud data by a pre-built target detection model, and the dynamic obstacle is tracked according to the target tracking algorithm to obtain the motion information of the dynamic obstacle; Based on the position information of the static obstacles and the motion information of the dynamic obstacles, a local path for avoiding the static obstacles and the dynamic obstacles is planned in real time in the motion direction of the mobile unmanned vehicle through an improved artificial potential field method; according to the local path, a motion control instruction of the mobile unmanned vehicle is generated, and the motion control instruction includes a speed control instruction and a steering control instruction of the mobile unmanned vehicle, and controls the mobile unmanned vehicle to move along the local path until it reaches a preset target position; Based on the position information of the static obstacles and the motion information of the dynamic obstacles, a local path for avoiding the static obstacles and the dynamic obstacles is planned in real time in the motion direction of the mobile unmanned vehicle through an improved artificial potential field method, including: A static Gaussian potential field function is established for the position information of the static obstacle, and a dynamic Gaussian potential field function is established for the motion information of the dynamic obstacle, wherein the peak positions of the static Gaussian potential field function and the dynamic Gaussian potential field function are both determined by the position of the obstacle, the dynamic peak position of the dynamic Gaussian potential field function is offset according to the speed direction of the dynamic obstacle, and the offset of the dynamic peak position of the dynamic Gaussian potential field function is proportional to the product of the moving speed of the dynamic obstacle and a preset sliding time window; The dynamic Gaussian potential field functions of all dynamic obstacles are weightedly superimposed, and combined with the static Gaussian potential field function to obtain a composite Gaussian potential field function that describes the environment in which the mobile unmanned vehicle is located, wherein the weight coefficient of each obstacle in the composite Gaussian potential field function is adaptively adjusted according to its type, size, and distance from the mobile unmanned vehicle; Determining the function gradient of the composite Gaussian potential field function according to the current position of the mobile unmanned vehicle, wherein the function gradient includes a positive gradient direction and a negative gradient direction, and using the negative gradient direction as the target movement direction of the mobile unmanned vehicle, and the target movement direction guides the mobile unmanned vehicle to stay away from static obstacles and dynamic obstacles; Taking the negative gradient direction as the preferred direction, combining the kinematic constraints of the mobile unmanned vehicle and the obstacle boundary constraints, and using a parameterized curve fitting algorithm to generate a local path; The composite Gaussian potential field function is as follows: ； The static Gaussian potential field function is as follows: ； The dynamic Gaussian potential field function is as follows: ； in, U(p,t) express t The composite Gaussian potential field function at time , U s (p) represents the static Gaussian potential field function, U d (p,t) express t The dynamic Gaussian potential field function at time , represents the position vector of the mobile unmanned vehicle in the two-dimensional plane, represents the position vector of the static obstacle, represents the position vector of the dynamic obstacle, A s , A d represent the amplitude of the static Gaussian potential field function and the amplitude of the dynamic Gaussian potential field function respectively, , Respectively represent the covariance matrix of the static Gaussian potential field function and the dynamic Gaussian potential field function, w i Indicates i The weight coefficient corresponding to the dynamic Gaussian potential field function is n Indicates the number of dynamic obstacles.

2. The method according to claim 1, characterized in that Performing dynamic obstacle target detection on the dynamic point cloud data by using a pre-built target detection model includes: Processing the dynamic point cloud data based on a pre-built target detection model, wherein the target detection model extracts multi-scale point cloud features of the dynamic point cloud data through a PointNet++ backbone network; After the PointNet++ backbone network, a multi-scale interactive attention module is cascaded. The multi-scale interactive attention module contains multiple parallel attention units, which process point cloud features of different scales respectively. The point cloud features at each scale are mapped to a low-dimensional embedding space, and the correlation between the point cloud features at different positions is calculated to obtain an attention map. The attention maps of different scales are adaptively aggregated through the fusion module of the target detection model to obtain multi-scale interactive features. The target detection model includes three parallel feature pyramid detection heads, which are responsible for the detection of targets of the first size, the second size and the third size respectively. Each feature pyramid detection head includes a center point prediction branch and a target size prediction branch. The center point prediction branch generates a center point heat map consistent with the multi-scale interactive feature resolution through convolution and upsampling layers. The target size prediction branch estimates the target three-dimensional bounding box size corresponding to each center point heat map through regression, wherein the first size, the second size and the third size are successively larger. The outputs of the three detection heads are post-processed through a non-maximum suppression module to obtain a multi-scale target detection result, and at least one of the category, position, size and orientation information of the dynamic obstacle in the dynamic point cloud data is output.

3. The method according to claim 2, characterized in that Tracking the dynamic obstacle according to the target tracking algorithm to obtain the motion information of the dynamic obstacle includes: The multi-scale target detection results are input into a multi-target tracking framework based on Kalman filtering, wherein the multi-target tracking framework adds an acceleration component to the target state space, and sets a state transfer matrix and a process noise covariance matrix in the target state space, wherein the state transfer matrix and the process noise covariance matrix contain prior statistical information of the target position, velocity and acceleration, and are used to characterize the nonlinear motion characteristics of the target; A two-stage cascade matching strategy is used to associate the multi-scale target detection result at the current moment with the previously acquired tracking trajectory. The two-stage cascade matching strategy includes a preliminary screening based on the cosine similarity of the appearance features and a fine screening based on the IoU geometric metric. A set of candidate matching pairs is obtained through the preliminary screening, and a fine screening is performed on this basis to obtain the final matching result. Based on the final matching results and the target prior motion model, the state parameters of each dynamic obstacle, including position, velocity and acceleration, are estimated through the prediction and update process of the extended Kalman filter, and a continuous and smooth tracking trajectory is generated.

4. The method according to claim 3, characterized in that: After obtaining the final matching result, the method further includes performing motion compensation on the mobile unmanned vehicle: The measurement information of the on-board IMU and wheel speed meter of the mobile unmanned vehicle is fused to estimate the motion state of the vehicle, and the transformation relationship between the vehicle coordinate system and the world coordinate system is recursively calculated according to the motion state of the vehicle. The prediction result and the observation result of the tracker of the mobile unmanned vehicle are transformed by using the transformation relationship, and the prediction result and the observation result are unified in the world coordinate system to realize the motion compensation of the vehicle.

5. An obstacle map marking system in combination with an unmanned vehicle, used to implement the method of any one of claims 1 to 4, characterized in that: include: The first unit is used to determine the moving direction and moving path of the mobile unmanned vehicle according to the current location information of the mobile unmanned vehicle and the preset target location information; through the laser radar scanning device installed on the mobile unmanned vehicle, the surrounding environment of the mobile unmanned vehicle is scanned in real time to obtain point cloud data reflecting the location information of static obstacles and the instantaneous location information of dynamic obstacles in the indoor environment; The second unit is used to cluster the point cloud data by a clustering segmentation algorithm, extract static point cloud data reflecting static obstacles and dynamic point cloud data reflecting dynamic obstacles from the point cloud data, and determine the position information of the static obstacles according to the clustered static point cloud data; perform dynamic obstacle target detection on the dynamic point cloud data by a pre-built target detection model, and track the dynamic obstacles according to the target tracking algorithm to obtain the motion information of the dynamic obstacles; The third unit is used to plan a local path that avoids the static obstacles and dynamic obstacles in real time in the movement direction of the mobile unmanned vehicle based on the position information of the static obstacles and the motion information of the dynamic obstacles through an improved artificial potential field method; according to the local path, generate motion control instructions for the mobile unmanned vehicle, the motion control instructions including speed control instructions and steering control instructions for the mobile unmanned vehicle, and control the mobile unmanned vehicle to move along the local path until it reaches a preset target position.

6. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images