Target identification tracking method and device for specific person based on solid-state laser radar
By employing point cloud clustering, bounding box fitting, and Kalman filtering algorithms, the problem of insufficient accuracy and real-time performance of LiDAR in target recognition and tracking under CPU environments was solved, achieving stable target recognition and tracking under noise interference and dynamic environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-19
- Publication Date
- 2026-03-20
AI Technical Summary
Existing lidar target recognition and tracking technologies in pure CPU environments suffer from insufficient accuracy and real-time performance. In particular, performance degradation is prone to occur in noisy and dynamic environments, leading to misidentification and target loss.
By employing point cloud clustering, bounding box fitting and filtering, and multi-target tracking management algorithms, combined with Kalman filtering, and confirming targets through data filtering, ground segmentation, Euclidean clustering, and 3D detection boxes, stable identification and tracking are achieved.
High-precision and stable target recognition and tracking were achieved under low computing resource conditions, reducing false recognition and target loss, and improving the robustness of the system.
Smart Images
Figure CN121703832A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of lidar technology, and in particular to a method and apparatus for target recognition and tracking of specific individuals based on solid-state lidar. Background Technology
[0002] As the demand for LiDAR in target recognition and tracking in complex scenarios continues to increase, related technologies are also developing rapidly. The need for high-precision, real-time target tracking of "people" under low computing power conditions is becoming increasingly urgent. Current target recognition and tracking technologies place high demands on computing resources in terms of processing speed and real-time performance. LiDAR-based solutions often rely on GPUs or high-performance computing devices, which have gained widespread application in autonomous driving, security monitoring, and personnel identification due to their high measurement accuracy, strong anti-interference capabilities, and good environmental adaptability. However, the performance of LiDAR systems in a pure CPU environment is limited, resulting in shortcomings in the accuracy and real-time performance of target recognition and tracking, especially in noisy and dynamic environments where performance degradation is more likely.
[0003] In dynamic environments, LiDAR data is often affected by noise interference. This noise may originate from reflected signals from other objects in the environment or errors in the sensor itself, resulting in a large number of outliers in the point cloud data. These outliers can affect the accuracy of human identification and tracking, causing misidentification or even target loss, and in severe cases, may cause the system to malfunction. Therefore, how to filter out noise, accurately identify human targets, and achieve stable tracking in complex scenes has become a key research direction for current LiDAR target recognition and tracking systems. Summary of the Invention
[0004] To address the aforementioned problems in the existing technology, this application provides a method and apparatus for target identification and tracking of specific individuals based on solid-state lidar. The aim is to achieve stable identification and tracking of target individuals by utilizing point cloud clustering, bounding box fitting and filtering, and multi-target tracking management algorithms to complete stable identification and tracking of target individuals.
[0005] Firstly, this application provides a method for target recognition and tracking of specific individuals based on solid-state lidar, used for identifying and tracking targets categorized as people. The method includes: S1. Perform data filtering and preprocessing on the laser point cloud data, and then perform ground segmentation processing on the laser point cloud. S2. Reduce the dimensionality of the point cloud data after ground segmentation from three-dimensional space to two-dimensional space. Then, perform Euclidean clustering based on the spatial distance of the point cloud data. Finally, store the clustering results according to the index array, with each clustering result corresponding to a target object. S3. After identifying the point cloud clusters of the target person, extract the 3D detection box of the point cloud clusters, and confirm whether the target is a specific person based on the size and shape features of the 3D detection box. S4. A multi-target tracking algorithm based on Kalman filtering predicts and updates the motion state of the target person in order to achieve stable tracking of the target person.
[0006] Optionally, the data filtering and preprocessing of the laser point cloud data includes: Distance filtering is performed on the laser point cloud data to remove invalid points and abnormally far points by setting an effective distance range; Then, statistical filtering methods are used to remove outliers. By setting the nearest neighbor number or mean error threshold, isolated points and noise points are effectively eliminated.
[0007] Optionally, the ground segmentation process for the laser point cloud includes: The point cloud is segmented along both the angle and the radial direction. The point with the lowest height in each radial segment is extracted as the Min-Z feature point for that segment. The Min-Z point for each segment is determined as follows: ; Then, a linear fit is performed on the Min-Z point set of each angle segment to construct a ground model. The ground line obtained by the fit is used to measure the vertical distance from each point to the ground line and compare it with the set error threshold. Points below the threshold are regarded as ground points, and those exceeding the threshold are judged as obstacle points.
[0008] Optionally, the step of performing Euclidean clustering based on the spatial distance of the point cloud data, and then storing the clustering results according to an index array, with each clustering result corresponding to a target object, specifically includes: The KdTree search method was used to search the 2D point cloud data. Then, the Euclidean clustering algorithm was used to cluster the non-terrestrial point clouds based on spatial distance, grouping point clouds belonging to the same object into a single cluster, with a set distance threshold. Minimum and maximum number of points in a cluster , To obtain clustering results The clustering results are stored in a local index array, with each clustering result corresponding to a target object; Finally, the clustered point cloud clusters were analyzed. By setting height and volume feature thresholds, the size of each cluster was analyzed. The size feature of the clusters was... ,in These represent the length, width, and height of the clusters, respectively. Based on a preset target person size threshold, clusters that meet the size characteristics are selected. .
[0009] Optionally, after identifying the point cloud clusters of the target person, the step of extracting the 3D detection bounding boxes of the point cloud clusters and confirming whether the target is a specific person based on the size and shape features of the 3D detection bounding boxes specifically includes: After identifying the point cloud clusters of the target person, the Graham scanning algorithm is used to calculate the convex hull of the point cloud clusters based on the two-dimensional projection of the point cloud data, and the vertex coordinates of the bounding box are obtained. Calculate the orientation angles corresponding to the longest m line segments in the contour based on the vertex coordinates of the bounding box. Then select the bounding box with the smallest SCD as the best bounding box and calculate the length, width, height and orientation angle of the bounding box. Based on the preset target person size, a detection box that matches the size characteristics is selected. Then, the size and shape characteristics of the 3D detection box are used to confirm whether the target is a specific person.
[0010] Optionally, the step of calculating the orientation angles corresponding to the longest m line segments in the contour based on the vertex coordinates of the bounding box, and then selecting the bounding box with the smallest SCD as the optimal bounding box, and calculating the length, width, height and orientation angles of the bounding box, specifically includes: Based on the vertex coordinates of the bounding box Calculate each line segment length and direction angle ,in, By selecting the four longest line segments, we obtain the set of direction angles. ; According to the set of direction angles For each orientation angle, fit a 2D bounding box with orientation, calculate the length, width, and height of the bounding box, and calculate the feature distance and SCD.
[0011] Optionally, the calculation of feature distance and SCD includes: Let the direction angle be... straight line The intercepts on the horizontal and vertical axes are respectively Then, all points within the bounding box Projected onto a straight line Projected coordinates on The formula for calculating the projected coordinates is as follows: ; ; From this, we can derive the minimum and maximum values of the coordinates of all projected points. Repeat this process, projecting all points within the bounding box onto a direction angle of... straight line Up, get ; The vertices of the rectangular bounding box are calculated based on the nearest and farthest points. ; ; The final value of SCD is It is the sum of the minimum distances from all points within the point cloud cluster to the four sides of the rectangular bounding box.
[0012] Optionally, confirming whether the target is a specific person based on the size and shape features of the 3D detection box includes: Calculate the dimensional features of the detection frame based on its size and shape characteristics. ,when Greater than the threshold set for this category When the target is a specific person, the weight function is designed as follows: .
[0013] Optionally, the Kalman filter-based multi-target tracking algorithm predicts and updates the motion state of the target person to achieve stable tracking, specifically including: Initialize the state vector of the Kalman filter Initialize the covariance matrix for The initial position of the target is... and initial velocity ; Let the time interval be Δt, then the state transition matrix F, the process noise covariance matrix Q, and the measurement noise covariance matrix R are respectively: ; ; ; According to the public notice Predict the state vector and covariance matrix, where, ; Based on the actual detected target location, update the state vector and covariance matrix of the Kalman filter. Let the measurement matrix be... Calculate Kalman gain And update the corresponding state vector and covariance matrix.
[0014] Secondly, this application also provides a target recognition and tracking device for specific persons based on solid-state lidar, used to identify and track targets of the category of people. The target recognition and tracking device for specific persons based on solid-state lidar is configured to perform the above-mentioned target recognition and tracking method for specific persons based on solid-state lidar.
[0015] The beneficial technical effects of this application are as follows: This application provides a method and apparatus for target recognition and tracking of specific individuals based on solid-state lidar. Addressing the problem of unstable target recognition and tracking under pure CPU computing conditions, this application achieves efficient and stable target recognition and tracking by combining lidar point cloud data clustering and Kalman filtering algorithms. For the interference of ground points and outliers on point cloud clustering, ground segmentation and data filtering are used to effectively remove ground points and outliers, improving the stability of point cloud clustering. For the identification of individuals, Euclidean clustering and weighted function filtering are used to accurately identify potential target objects and select targets that conform to human body size characteristics. To address the issues of target loss and jumping during tracking, a target association and tracking management mechanism is designed. Through target association and Kalman filtering algorithms, stable target tracking is achieved, reducing tracking failures. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A flowchart illustrating a target recognition and tracking method for a specific person based on solid-state lidar provided in this application; Figure 2 A flowchart illustrating a target recognition and tracking method for a specific person based on solid-state lidar provided in this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0020] It should be understood that in this application, "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product or device.
[0021] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0022] refer to Figure 1 As shown, this invention provides a method for target identification and tracking of specific individuals based on solid-state LiDAR, used for identifying and tracking targets categorized as people. The core of this method lies in efficient and accurate point cloud data processing technology, employing a point cloud clustering algorithm for target identification, combined with Kalman filtering for stable tracking. This method is applicable to target identification and tracking in complex environments, including both static and dynamic environments. Specific application scenarios include, but are not limited to: security monitoring, personnel identification, detection of pedestrians and obstacles by autonomous vehicles, and following of specific individuals by autonomous vehicles.
[0023] refer to Figure 1 As shown, this embodiment of the invention provides a method for target recognition and tracking of a specific person based on solid-state lidar, including the following steps: S1. Perform data filtering and preprocessing on the laser point cloud data, and then perform ground segmentation processing on the laser point cloud.
[0024] This invention employs data filtering and ground segmentation to accurately distinguish between ground points and obstacle points, adapting to the point cloud processing needs in complex environments. It utilizes a highly efficient algorithm combining distance filtering, outlier removal, angular radial segmentation, and local linear fitting to achieve high-precision ground segmentation with low computational resources.
[0025] This invention provides a target recognition and tracking method for specific individuals based on solid-state lidar. The method performs distance filtering on lidar point cloud data, removing invalid and abnormally far points by setting an effective distance range. Then, statistical filtering is used to remove outliers, effectively eliminating isolated and noisy points by setting the nearest neighbor count or a mean error threshold. Lidar-acquired point cloud data often contains a large amount of noise and outliers; therefore, data filtering is necessary to improve the accuracy and efficiency of subsequent processing. First, distance filtering is performed to remove invalid and abnormally far points by setting an effective distance range, ensuring that only points within the specified range are retained, thereby improving computational efficiency. Next, statistical filtering and other methods are used to remove outliers, effectively eliminating isolated and noisy points and reducing their interference with ground segmentation by setting the nearest neighbor count (e.g., 5) or a mean error threshold (e.g., 0.2 meters).
[0026] The point cloud is segmented along both the angle and the radial direction. The point with the lowest height in each radial segment is extracted as the Min-Z feature point for that segment. The Min-Z point for each segment is determined as follows: Then, a linear fit is performed on the Min-Z point set of each angle segment to construct a ground model. The ground line obtained by fitting is used to measure the vertical distance from each point to the ground line and compare it with the set error threshold. Points below the threshold are regarded as ground points, and those exceeding the threshold are judged as obstacle points.
[0027] After data filtering, the point cloud data is segmented to identify ground points and non-ground points. In this step, the point cloud is first segmented angularly and radially, and the point with the lowest elevation in each radial segment is extracted as the feature point (Min-Z point) for that segment. This significantly reduces computation and highlights ground features. The Min-Z point for each segment is determined as follows: Next, a linear fit is performed on the Min-Z point set for each angle segment to construct the ground model.
[0028] Local line segment fitting and ground determination: Least square linear fitting is performed on the Min-Z point set for each angle segment to obtain the ground line segment model. The fitted line equation is: Where a and b are the linear parameters of the fitted graph, and d is the distance. The ground line obtained through fitting is used to measure the vertical distance from each point to the line and compare it with a set error threshold. Points below the threshold are considered ground points, while those exceeding the threshold are identified as obstacle points. This method achieves accurate separation of the ground and obstacles.
[0029] For example, in the data filtering stage, the effective distance range of the point cloud is set to 1 to 50 meters to remove invalid points outside this range. Furthermore, statistical filtering is applied to remove outliers; for instance, a nearest neighbor threshold of 5 is set for each point, and a mean error threshold of 0.2 meters is set to effectively clean up isolated and noisy points, ensuring the accuracy and representativeness of the remaining data. In the ground segmentation process, the point cloud is divided into 180 segments by angle, with an angle step of approximately 2 degrees per segment; and 100 segments radially, with a distance step of approximately 0.5 meters per segment. Then, the lowest elevation point (Min-Z point) is extracted within each radial segment, and a linear fit is performed on the Min-Z point set for each angle segment. Finally, points with a vertical distance of less than 0.1 meters to the fitted line are identified as ground points, and the rest are considered obstacle points. This method significantly improves the accuracy of separating the ground from obstacles.
[0030] S2. Reduce the dimensionality of the point cloud data after ground segmentation from three-dimensional space to two-dimensional space. Then, perform Euclidean clustering based on the spatial distance of the point cloud data. Finally, store the clustering results in an index array, with each clustering result corresponding to a target object.
[0031] The KdTree search method was used to search the 2D point cloud data. Then, the Euclidean clustering algorithm was used to cluster the non-terrestrial point clouds based on spatial distance, grouping point clouds belonging to the same object into a single cluster, with a set distance threshold. Minimum and maximum number of points in a cluster , To obtain clustering results The clustering results are stored in a local index array, with each cluster corresponding to a target object. Finally, the clustered point cloud clusters are analyzed. By setting height and volume feature thresholds, the size of each cluster is analyzed. The size feature of the cluster is... ,in These represent the length, width, and height of the clusters, respectively. Based on a preset target person size threshold, clusters that meet the size characteristics are selected. .
[0032] Specifically: 1) Point cloud dimensionality reduction processing: Point cloud data is reduced from 3D space to 2D space by ignoring height information to reduce computational complexity and facilitate subsequent clustering processing, resulting in 2D point cloud data. ,in, The z-coordinate of each point in the point cloud data is set to 0 to obtain two-dimensional point cloud data; then the KdTree search method is used to search the two-dimensional point cloud data to improve clustering efficiency.
[0033] 2) Euclidean clustering based on KdTree: The KdTree search method is used to search 2D point cloud data to improve clustering efficiency. The Euclidean clustering algorithm is used to cluster non-terrestrial point clouds based on spatial distance, grouping point clouds belonging to the same object into a single cluster. A distance threshold ϵ and a minimum and maximum number of points per cluster are set. To obtain clustering results The clustering results are stored in a local index array, with each clustering result corresponding to a target object.
[0034] Based on the spatial distance of point cloud data, point cloud data belonging to the same target object are grouped together. The Euclidean Cluster algorithm is used to cluster non-terrestrial point clouds, grouping point clouds belonging to the same object into a single cluster. During the Euclidean Cluster algorithm clustering process, a clustering distance threshold of 0.1m is set, and the minimum and maximum number of cluster points are set to 100 and 8000 respectively. This effectively identifies objects of different shapes and sizes, ensuring the reasonableness of the clustering results.
[0035] 3) Clustering result screening The clustering results are stored in a local index array, with each cluster corresponding to a target object. Finally, the clustered point cloud clusters are analyzed. By setting feature thresholds such as height and volume, point cloud clusters that may contain the size features of the target person are selected. Size analysis is performed on each cluster, assuming the size feature of the cluster is... ,in These represent the length, width, and height of the cluster, respectively. Based on a preset target person size threshold, clusters matching the size characteristics are selected. .
[0036] The clustering results are stored in a local index array, with each cluster corresponding to a target object. Finally, the clustered point cloud clusters are analyzed, and the minimum and maximum cluster sizes are adjusted to 0.2m and 15m respectively to remove point cloud clusters that are too small or too large from the scene.
[0037] S3. After identifying the point cloud clusters of the target person, extract the 3D detection box of the point cloud clusters, and confirm whether the target is a specific person based on the size and shape features of the 3D detection box.
[0038] After identifying the point cloud clusters of the target person, the Graham scanning algorithm is used to calculate the convex hull of the point cloud clusters based on the two-dimensional projection of the point cloud data, and the vertex coordinates of the bounding box are obtained. Calculate the orientation angles corresponding to the longest m line segments in the contour based on the vertex coordinates of the bounding box. Then select the bounding box with the smallest SCD as the best bounding box and calculate the length, width, height and orientation angle of the bounding box. Based on the preset target person size, a detection box that matches the size characteristics is selected. Then, the size and shape characteristics of the 3D detection box are used to confirm whether the target is a specific person.
[0039] Extracting 3D bounding boxes from point cloud clusters, specifically including: 1) 2D Convex Hull Calculation: The point cloud data is reduced from 3D space to 2D space, ignoring height information to reduce computational complexity and facilitate subsequent bounding box fitting. During the convex hull calculation, the Graham scan algorithm is used to calculate the convex hull of the point cloud cluster based on the 2D projection of the point cloud data, obtaining the vertex coordinates of the bounding box. A reference point p0 is selected, usually the point with the smallest y-coordinate in the point set. If multiple points have the same y-coordinate, the point with the smallest x-coordinate is selected. All points are then sorted according to their polar angle with the reference point p0. Starting from the reference point p0, the point set is traversed in sorted order. The first two points p0 and p1 are pushed onto a stack. For each subsequent point pi, it is checked whether the top two points of the stack and the current point pi form a left turn (counter-clockwise). If it's a left turn, push pi onto the stack; otherwise, pop the top of the stack until a left turn is formed. The final point in the stack is the vertex of the convex hull.
[0040] 2) Bounding box fitting: Based on the coordinates of the convex hull vertices, calculate the direction angles corresponding to the m longest line segments in the contour. Let the coordinates of the convex hull vertices be... ,in, Then calculate each line segment. length and direction angle Then there is ; ; Select the four longest line segments to obtain the set of direction angles. According to the set of direction angles For each orientation angle, a 2D bounding box with orientation is fitted. The length, width, and height of the bounding box are calculated, and the sum of feature distances (SCD) is calculated. The bounding box with the smallest SCD is selected as the optimal bounding box. The calculation process of SCD is as follows: Let the direction angle be... straight line The intercepts on the horizontal and vertical axes are respectively Then, all points within the bounding box Projected onto a straight line Projected coordinates on The formula for calculating the projected coordinates is as follows: ; ; Projected coordinates onto this line From this, the minimum and maximum values of the coordinates of all projected points can be obtained. Similarly, project all points within the bounding box onto a direction angle of... straight line ,get The vertices of the rectangular bounding box are calculated based on the nearest and farthest points: ; ; The final value of SCD is ; This is the sum of the minimum distances from all points within the point cloud cluster to the four sides of the rectangular bounding box. Finally, the bounding box with the smallest SCD is selected as the optimal bounding box, and its length, width, height, and orientation angle are calculated.
[0041] 3) Target Classification: Based on the preset target human body dimensions, select detection boxes that match the size characteristics. Then, analyze the shape of the detection boxes to confirm whether they conform to human body shape characteristics.
[0042] The weighting function is designed as follows: .
[0043] calculate ,when When the value exceeds the threshold set for that category, the target is considered to belong to the corresponding category. Let be a Gaussian distribution function, where The standard deviation of the Gaussian distribution is set. The set distance threshold is generally related to... Equal; when The time is a linear function, and smoothness is guaranteed at the connection points.
[0044] Final calculation ,when If the value exceeds the threshold set for that category, the target is considered to belong to the corresponding category. Finally, the identified target object information is published to the ROS system for subsequent processing and display.
[0045] S4. A multi-target tracking algorithm based on Kalman filtering predicts and updates the motion state of the target person in order to achieve stable tracking of the target person.
[0046] Initialize the state vector of the Kalman filter Initialize the covariance matrix for The initial position of the target is... and initial velocity ; Let the time interval be Δt, then the state transition matrix F, the process noise covariance matrix Q, and the measurement noise covariance matrix R are respectively: ; ; ; According to the public notice Predict the state vector and covariance matrix, where, ; Based on the actual detected target location, update the state vector and covariance matrix of the Kalman filter. Let the measurement matrix be... Calculate Kalman gain And update the corresponding state vector and covariance matrix.
[0047] The multi-target tracking algorithm based on Kalman filtering predicts and updates the motion state of a target, specifically including: 1) Kalman Filtering and State Prediction: Initialize the state vector and covariance matrix of the Kalman filter, and set the initial position of the target. and initial velocity The initial state vector of the Kalman filter is: Initialize the covariance matrix for: ; Let the time interval be Δt, then the state transition matrix F, the process noise covariance matrix Q, and the measurement noise covariance matrix R are respectively: ; ; ; According to the public notice Predict the state vector and covariance matrix, where, .
[0048] 2) Status update and target association Based on the actual detected target location, update the state vector and covariance matrix of the Kalman filter. Let the measurement matrix be... Calculate Kalman gain And update the corresponding state vector and covariance matrix.
[0049] During each frame of data processing, detected targets are associated with tracked targets, and the target states are updated. Let the set of detected targets be... The set of tracked targets is Calculate the distance between each detected target and the tracked target, and find the matching pair with the smallest distance. To establish a relationship, treat the related objects as the same target.
[0050] A target loss threshold of L is set. If a tracked target is not detected for L consecutive frames, it is considered lost and removed from the tracked target set. If a new target is detected and does not match any tracked target, it is added to a temporary target set for further confirmation. If a temporary target is detected in two consecutive frames, it is added to the tracked target set.
[0051] Manage the lifecycle of tracked targets, handling target loss and reappearance. Calculate the distance of each tracked target from the origin: Let the set of tracked targets be... The current target being followed has the ID C, and the next target to be followed has the ID N. Calculate the distance between each tracked target and the origin. The coordinates are The distance from the origin is Find the nearest object. Let the ID of the nearest object be Nc, and the distance be... .
[0052] The update is based on the current state of the target being followed, and the switching logic is as follows: If the ID of the current target being followed is -1 (i.e., no target is being followed), then the nearest object is taken as the current target being followed, and let... If the ID of the currently tracked target is not -1, check if the currently followed target has disappeared. If the currently followed target has disappeared, then the next followed target or the nearest object is set as the current followed target. If the currently followed target has not disappeared, but a new, closer object is found, then the ID of the next followed target is set. and reset the switching time. During the processing of each frame of data, it checks whether the switching time has arrived. If the switching time has arrived, it switches to the next target to follow.
[0053] This application provides a method and apparatus for target recognition and tracking of specific individuals based on solid-state lidar. Addressing the problem of unstable target recognition and tracking under pure CPU computing conditions, this application achieves efficient and stable target recognition and tracking by combining lidar point cloud data clustering and Kalman filtering algorithms. For the interference of ground points and outliers on point cloud clustering, ground segmentation and data filtering are used to effectively remove ground points and outliers, improving the stability of point cloud clustering. For the identification of individuals, Euclidean clustering and weighted function filtering are used to accurately identify potential target objects and select targets that conform to human body size characteristics. To address the issues of target loss and jumping during tracking, a target association and tracking management mechanism is designed. Through target association and Kalman filtering algorithms, stable target tracking is achieved, reducing tracking failures.
[0054] Specifically, in the experimental platform established in this invention, a Livox HAP solid-state lidar is used. Data filtering, point cloud clustering, and target tracking modules are simultaneously activated on the ROS platform. Following the aforementioned parameter settings, 1-5 people are tracked outdoors, and the tracking process is interfered with. The designed interference process includes: Some people quickly pass by the tracked target from near to far; some people quickly enter and exit the lidar's field of view at close range; some people replace the tracked target at close range; when someone is close, the tracked target suddenly disappears; when someone is far away, the tracked target disappears; some people stay far away from the tracked target; two or more people exchange distances near the tracked target and frequently disappear.
[0055] Even during interference, the success rate of identifying specific targets still reaches over 95%. This indicates that through the above steps, the present invention can effectively achieve stable identification and tracking of target individuals, providing accurate location, size, and speed information. The target identification and tracking method for specific individuals based on solid-state LiDAR provided in this application, by combining LiDAR point cloud data clustering and multi-target tracking algorithms, achieves a tracking success rate of over 97% in pre-recorded data packets (including parking lots, underground passages, and narrow lanes) using the technical solution of this invention.
[0056] Based on the same inventive concept, this application also provides a target recognition and tracking device for a specific person based on solid-state lidar, which can execute the processing flow provided in an embodiment of a target recognition and tracking method for a specific person based on solid-state lidar.
[0057] This application provides a target recognition and tracking device for a specific person based on solid-state lidar. This device can be used to execute any of the target recognition and tracking methods for a specific person based on solid-state lidar described in the above embodiments. The implementation principle and technical effects are similar and will not be repeated here. The device provided in this application can be specifically used to execute the solutions provided in the corresponding method embodiments; the specific functions and achievable technical effects will not be repeated here.
[0058] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. Additionally, these modules can be fully or partially integrated together, or implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through the integrated logic circuits in the hardware of the processor element or through software instructions.
[0059] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 As shown, the electronic device may include: a processor 21, a memory 22, and computer program instructions stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program instructions, it implements a target recognition and tracking method for a specific person based on solid-state lidar provided in any of the foregoing embodiments.
[0060] Optionally, the various components of the electronic device can be connected via a system bus.
[0061] The memory 22 can be a separate memory unit or a memory unit integrated into the processor. The number of processors can be one or more.
[0062] Optionally, the electronic device may also include a communication interface for interacting with other devices.
[0063] It should be understood that the processor 21 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.
[0064] The system bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. Memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0065] All or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable memory. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof.
[0066] The electronic device provided in this application embodiment can be used to execute a target recognition and tracking method for a specific person based on solid-state lidar provided in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.
[0067] This application provides a computer-readable storage medium storing computer-executable instructions. When these instructions are executed on a computer, the computer performs the aforementioned target recognition and tracking method for a specific person based on solid-state LiDAR.
[0068] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.
[0069] Optionally, a readable storage medium can be coupled to the processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.
[0070] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the above-described target recognition and tracking method for specific persons based on solid-state lidar.
[0071] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A method for target recognition and tracking of specific individuals based on solid-state lidar, used for identifying and tracking targets categorized as people, characterized in that... The target recognition and tracking method for specific individuals based on solid-state lidar includes: S1. Perform data filtering and preprocessing on the laser point cloud data, and then perform ground segmentation processing on the laser point cloud. S2. Reduce the dimensionality of the point cloud data after ground segmentation from three-dimensional space to two-dimensional space. Then, perform Euclidean clustering based on the spatial distance of the point cloud data. Finally, store the clustering results according to the index array, with each clustering result corresponding to a target object. S3. After identifying the point cloud clusters of the target person, extract the 3D detection box of the point cloud clusters, and confirm whether the target is a specific person based on the size and shape features of the 3D detection box. S4. A multi-target tracking algorithm based on Kalman filtering predicts and updates the motion state of the target person in order to achieve stable tracking of the target person.
2. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 1, characterized in that, The data filtering and preprocessing of the laser point cloud data includes: Distance filtering is performed on the laser point cloud data to remove invalid points and abnormally far points by setting an effective distance range; Then, statistical filtering methods are used to remove outliers. By setting the nearest neighbor number or mean error threshold, isolated points and noise points are effectively eliminated.
3. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 1, characterized in that, The ground segmentation process for the laser point cloud includes: The point cloud is segmented along both the angle and the radial direction. The point with the lowest height in each radial segment is extracted as the Min-Z feature point for that segment. The Min-Z point for each segment is determined as follows: ； Then, a linear fit is performed on the Min-Z point set of each angle segment to construct a ground model. The ground line obtained by the fit is used to measure the vertical distance from each point to the ground line and compare it with the set error threshold. Points below the threshold are regarded as ground points, and those exceeding the threshold are judged as obstacle points.
4. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 1, characterized in that, The step involves performing Euclidean clustering based on the spatial distance of the point cloud data, followed by storing the clustering results in an index array, with each clustering result corresponding to a target object. Specifically, this includes: The KdTree search method was used to search the 2D point cloud data. Then, the Euclidean clustering algorithm was used to cluster the non-terrestrial point clouds based on spatial distance, grouping point clouds belonging to the same object into a single cluster, with a set distance threshold. Minimum and maximum number of points in a cluster , To obtain clustering results The clustering results are stored in a local index array, with each clustering result corresponding to a target object; Finally, the clustered point cloud clusters were analyzed. By setting height and volume feature thresholds, the size of each cluster was analyzed. The size feature of the clusters was... ,in These represent the length, width, and height of the clusters, respectively. Based on a preset target person size threshold, clusters that meet the size characteristics are selected. .
5. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 1, characterized in that, After identifying the point cloud clusters of the target person, the process of extracting the 3D detection bounding boxes of the point cloud clusters and confirming whether the target is a specific person based on the size and shape features of the 3D detection bounding boxes specifically includes: After identifying the point cloud clusters of the target person, the Graham scanning algorithm is used to calculate the convex hull of the point cloud clusters based on the two-dimensional projection of the point cloud data, and the vertex coordinates of the bounding box are obtained. Calculate the orientation angles corresponding to the longest m line segments in the contour based on the vertex coordinates of the bounding box. Then select the bounding box with the smallest SCD as the best bounding box and calculate the length, width, height and orientation angle of the bounding box. Based on the preset target person size, a detection box that matches the size characteristics is selected. Then, the size and shape characteristics of the 3D detection box are used to confirm whether the target is a specific person.
6. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 5, characterized in that, The process involves calculating the orientation angles corresponding to the longest m line segments in the contour based on the vertex coordinates of the bounding box, then selecting the bounding box with the smallest SCD as the optimal bounding box, and calculating the length, width, height, and orientation angles of the bounding box. Specifically, this includes: Based on the vertex coordinates of the bounding box Calculate each line segment length and direction angle ,in, By selecting the four longest line segments, we obtain the set of direction angles. ; According to the set of direction angles For each orientation angle, fit a 2D bounding box with orientation, calculate the length, width, and height of the bounding box, and calculate the feature distance and SCD.
7. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 6, characterized in that, The calculation of feature distance and SCD includes: Let the direction angle be... straight line The intercepts on the horizontal and vertical axes are respectively Then, all points within the bounding box Projected onto a straight line Projected coordinates on The formula for calculating the projected coordinates is as follows: ； ； From this, we can derive the minimum and maximum values of the coordinates of all projected points. Repeat this process, projecting all points within the bounding box onto a direction angle of... straight line Up, get ; The vertices of the rectangular bounding box are calculated based on the nearest and farthest points. ； ； The final value of SCD is It is the sum of the minimum distances from all points within the point cloud cluster to the four sides of the rectangular bounding box.
8. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 6, characterized in that, The step of confirming whether a target is a specific person based on the size and shape features of the 3D detection frame includes: Calculate the dimensional features of the detection frame based on its size and shape characteristics. ,when Greater than the threshold set for this category When the target is a specific person, the weight function is designed as follows: 。 9. The target recognition and tracking method for a specific person based on solid-state lidar according to claim 1, characterized in that, The Kalman filter-based multi-target tracking algorithm predicts and updates the motion state of the target to achieve stable tracking of the target, specifically including: Initialize the state vector of the Kalman filter Initialize the covariance matrix for The initial position of the target is... and initial velocity ; Let the time interval be Δt, then the state transition matrix F, the process noise covariance matrix Q, and the measurement noise covariance matrix R are respectively: ； ； ； According to the public notice Predict the state vector and covariance matrix, where, ; Based on the actual detected target location, update the state vector and covariance matrix of the Kalman filter. Let the measurement matrix be... Calculate Kalman gain And update the corresponding state vector and covariance matrix.
10. A target recognition and tracking device for specific persons based on solid-state lidar, used for recognizing and tracking targets classified as people, characterized in that, The target recognition and tracking device for a specific person based on solid-state lidar is configured to perform the target recognition and tracking method for a specific person based on solid-state lidar as described in any one of claims 1 to 9.