Intelligent intersection target tracking method, device, equipment and medium

Abstract

This application discloses a smart intersection target tracking method, device, equipment, and medium, applied in the field of smart transportation technology. The method includes: when a disappearing target object is identified among various target objects, obtaining sparse point cloud clusters and newly formed point cloud clusters from the point cloud clusters acquired by lidar in subsequent frames, and determining the positions of the sparse point cloud clusters and the newly formed target clusters; predicting the initial positions of the disappearing target object in subsequent frames based on historical tracking data of the disappearing target object and lane line data of the target intersection; determining the target positions of the disappearing target object in subsequent frames based on the sparse point cloud positions, the newly formed target positions, and the initial positions; and reconstructing the target trajectory based on the target positions to obtain the target motion trajectory. By using sparse point cloud clusters and newly formed point cloud clusters as reference points for trajectory prediction, the computational load can be reduced, and the continuity and accuracy of target perception data at smart intersections can be improved.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and in particular to a method, device, equipment and medium for intelligent intersection target tracking. Background Technology

[0002] With the development of intelligent transportation, LiDAR is increasingly being used in vehicle perception scenarios at traffic intersections. By using point cloud data collected by LiDAR to perceive and track the trajectory of traffic targets, it can help traffic light control systems to more accurately control the changes in traffic lights based on information such as the position, type, number, and speed of vehicles on the road. It can also analyze the traffic efficiency, traffic load, and safety of intersections based on the trajectory information of targets, thereby enhancing the rationality of traffic planning. Furthermore, it can help traffic management departments to promptly detect and handle traffic accidents and congestion, thereby ensuring road traffic safety.

[0003] In existing technologies, due to the limited scanning range of lidar, blind spots exist in the lower area of ​​lidar. Moreover, due to the complexity of the surrounding environment at intersections, the lidar perception area is easily obstructed by green trees, road signs, large vehicles, etc., creating blind spots. These blind spots contain only scattered sparse point clusters, making it difficult to achieve target recognition. In addition, due to the many interference factors in target tracking, the target's identity identifier (entity identifier) ​​is prone to interruption or change during the target tracking process. When traffic is heavy, the identity identifier of one target may be assigned to another target near the blind spot.

[0004] To address the issue of target disappearance, the traditional approach is to predict the target's position two or three frames after it disappears by tracking its historical trajectory. This traditional tracking method is only suitable for temporarily compensating for a disappeared target by tracking and prediction during a short disappearance period. If the target is occluded for a long time or the blind spot is large, it will cause problems such as target tracking failure, interruption of target identification, or the target falling on other targets around the blind spot. This affects the continuity and accuracy of target perception data at smart intersections and interferes with traffic flow statistics, real-time tracking, and behavior analysis. Summary of the Invention

[0005] This application provides a method, apparatus, device, and medium for target tracking at smart intersections, to address the problem of poor continuity and accuracy of target perception data at smart intersections in the prior art. Specifically, the technical solution provided by this application is as follows:

[0006] On the one hand, this application provides a smart intersection target tracking method, including:

[0007] Based on the point cloud cluster of the target intersection collected by the lidar at the current frame, when tracking the target object of the target intersection within the point cloud perception range of the lidar, if it is determined that a target object with a missing identity tag appears within the point cloud perception range of the lidar, then the target object with a missing identity tag is identified as the missing target object.

[0008] Based on the historical tracking data before the disappearance of the missing target object's identification and the blind zone range of the lidar point cloud, the target time when the missing target object reappears within the lidar point cloud perception range is determined.

[0009] From the point cloud clusters collected by the lidar at each frame within the target time range based on the target time, sparse point cloud clusters within the point cloud blind zone and newly formed point cloud clusters at the boundary of the point cloud blind zone are obtained. Based on each sparse point cloud cluster at each frame, the positions of each sparse point cloud at each frame are determined, and based on each newly formed point cloud cluster at each frame, the positions of each newly formed target at each frame are determined. Among them, each newly formed point cloud cluster at each frame is a point cloud cluster that appears at the boundary of the point cloud blind zone at the frame.

[0010] Based on historical tracking data of the missing target object and map data of the target intersection, predict the initial position of the missing target object at each frame within the target time range with the target time as the reference.

[0011] Based on the sparse point cloud positions, newly generated target positions, and initial positions at each frame, determine the target positions of the disappearing target objects at each frame.

[0012] Based on the target positions of the vanished target object at each frame time, the target motion trajectory of the vanished target object is obtained by trajectory reconstruction.

[0013] On the other hand, this application provides a smart intersection target tracking device, comprising:

[0014] The disappearing target identification module is used to track target objects at the target intersection within the point cloud perception range of the LiDAR based on the point cloud clusters of the target intersection at the current frame. If it is determined that a target object with a disappearing identity tag appears within the point cloud perception range of the LiDAR, then the target object with the disappearing identity tag is identified as the disappearing target object.

[0015] The disappearing target location prediction module is used to determine the target time when the disappeared target object reappears within the point cloud perception range of the LiDAR, based on the historical tracking data before the disappearance of the disappeared target object and the blind zone range of the LiDAR point cloud. It extracts sparse point cloud clusters within the blind zone range and newly formed point cloud clusters at the boundaries of the blind zone from the point cloud clusters collected by the LiDAR in each frame within the target time range based on the target time. Based on each sparse point cloud cluster in each frame, it determines the position of each sparse point cloud cluster, and based on each newly formed point cloud cluster, it determines the position of each newly formed target in each frame. Each newly formed point cloud cluster in each frame is a cluster that appears at the boundary of the blind zone range in that frame. Based on the historical tracking data of the disappeared target object and the map data of the target intersection, it predicts the initial position of the disappeared target object in each frame within the target time range based on the target time. Based on the sparse point cloud positions, newly formed target positions, and initial positions in each frame, it determines the target position of the disappeared target object in each frame.

[0016] The vanishing target trajectory reconstruction module is used to reconstruct the trajectory of the vanishing target object based on its position at each frame time to obtain the target motion trajectory of the vanishing target object.

[0017] On the other hand, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned smart intersection target tracking method.

[0018] On the other hand, this application provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the above-described intelligent intersection target tracking method.

[0019] The beneficial effects of this application are as follows:

[0020] This application uses sparse point cloud clusters and newly formed point cloud clusters as reference points for location trajectory prediction to predict the target movement trajectory of a disappearing target object. This can improve the continuity and accuracy of target perception data at smart intersections while reducing the amount of computation.

[0021] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 This is a simplified flowchart illustrating the target tracking method at a smart intersection in this application.

[0024] Figure 2 This is a schematic diagram outlining the intelligent intersection target tracking method in the embodiments of this application;

[0025] Figure 3 This is a schematic diagram illustrating the specific process of the intelligent intersection target tracking method in the embodiments of this application;

[0026] Figure 4 This is a functional structure diagram of the smart intersection target tracking device in the embodiments of this application;

[0027] Figure 5 This is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and beneficial effects of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] To facilitate a better understanding of this application by those skilled in the art, the technical terms used in this application will be briefly introduced below.

[0030] TNT (Target-driven N Trajectory Prediction): A trajectory prediction model that defines the modality of each trajectory at the endpoint of each trajectory.

[0031] VectorNet: A hierarchical graph neural network. Compared to CNN, VectorNet models the relationships between instances more directly and allows distant instances to exchange information directly, enabling better extraction of features such as road information and motion trajectories.

[0032] Sparse point cloud clusters: Point cloud clusters in the point cloud blind zone of a lidar system, consisting of a small amount of point cloud data, which make it difficult to identify target objects.

[0033] Frenet coordinate system: The Frenet coordinate system is established based on a reference line, usually defined as the center line of the road, so that the horizontal and vertical coordinates are consistent with the road direction.

[0034] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein.

[0035] After introducing the technical terms used in this application, the application scenarios and design concepts of this application will be briefly introduced next.

[0036] When tracking target objects at a target intersection using point cloud clusters collected by LiDAR, blind spots or sparse point cloud areas may form within the LiDAR's point cloud perception range due to factors such as road vegetation, large vehicles obstructing the view, and environmental conditions. This can cause tracking interruptions and the disappearance of vehicles and other target objects. However, the actual vehicles will not disappear and will reappear at some point in the future. Traditional target tracking algorithms have a limited lifespan for tracking target objects. After a short prediction period, a disappeared target object will either disappear or be tracked onto another target object. When a disappeared target object leaves the point cloud blind spot or sparse point cloud area and is detected again, it will be treated as a new target object and given a completely new identity for tracking. This can lead to problems such as tracking and trajectory prediction errors, discontinuous and unreasonable identity identification, and duplicate road traffic flow.

[0037] To address this, this application utilizes intersection map data, contextual data, and newly generated or sparse point cloud clusters appearing in subsequent frames to perform multiple long-distance, long-duration trajectory predictions for the disappearing target object. The most reasonable predicted position is then selected and matched with the disappearing target object to ensure uninterrupted identification, thereby resolving the issues of target object tracking interruption and inaccuracy. Furthermore, this application improves the TNT trajectory prediction model by using sparse point cloud clusters (where target object identification is impossible) and newly generated point cloud clusters in subsequent frames as reference points for the TNT trajectory mode. This allows for a reasonable prediction of the target position and trajectory during the tracking interruption or disappearance period, optimizing the accuracy and reasonableness of the traditional TNT trajectory prediction model while reducing computational load.

[0038] After introducing the application scenarios and design concepts of this application, the technical solutions provided by this application will be described in detail below.

[0039] This application provides a smart intersection target tracking method, which can be applied to electronic devices such as computers and servers. For details, please refer to [link / reference]. Figure 1As shown, the general flow of the intelligent intersection target tracking method provided in this application embodiment is as follows:

[0040] Step 101: Based on the point cloud cluster of the target intersection collected by the LiDAR at the current frame time, when tracking the target object of the target intersection within the point cloud perception range of the LiDAR, if it is determined that a target object with a missing identity mark appears within the point cloud perception range of the LiDAR, then the target object with the missing identity mark is identified as the missing target object.

[0041] In this embodiment of the application, in order to predict the initial position of the disappearing target object, electronic devices such as computers and servers collect map data of the target intersection in advance, including but not limited to lane lines, zebra crossings, traffic signs, etc., and measure the effective point cloud perception range and fixed point cloud blind zone range of the LiDAR at the target intersection in advance. The edge of the effective point cloud perception range of the LiDAR is set as the target generation position, that is, when the target object enters the point cloud perception range of the LiDAR, it will be tracked as a new target.

[0042] Based on this, the LiDAR at the target intersection collects point cloud clusters of the target intersection in real time and uploads them to the electronic device. The electronic device uses the TNT trajectory prediction model to perform position tracking and trajectory prediction of the target object based on the pre-collected map data of the target intersection and the point cloud clusters of the target intersection collected by the LiDAR at each frame. For ease of calculation, in this embodiment, the target coordinate system is transformed from the Cartesian coordinate system to the Frenet coordinate system to perform position tracking and trajectory prediction of the target object perceived and identified by the LiDAR. That is, the sparse point cloud position, the position of the newly formed target, the initial position of the disappearing target object, and the target position mentioned later all belong to the position data in the Frenet coordinate system, and the following focuses on the TNT trajectory prediction. The model performs position tracking and trajectory prediction for disappearing target objects. The TNT trajectory prediction model mainly includes three stages when performing position tracking and trajectory prediction for disappearing target objects: disappearing target identification stage, disappearing target position prediction stage, and disappearing target trajectory reconstruction stage. In the disappearing target identification stage, based on the point cloud cluster of the target intersection collected by the lidar at the current frame, when tracking the target object at the target intersection within the lidar's point cloud perception range, if it is determined that a target object disappearing due to interruption or change of identity identification within the lidar's point cloud perception range is identified as a disappearing target object, the target object with interruption or change of identity identification is identified as a disappearing target object, and the historical tracking data such as the position, heading angle, and speed of the disappearing target object are recorded.

[0043] Step 102: Based on the historical tracking data of the disappeared target object before its disappearance and the blind zone range of the LiDAR point cloud, determine the target time when the disappeared target object reappears within the LiDAR point cloud perception range; from the point cloud clusters collected by the LiDAR in each frame within the target time range based on the target time, obtain the sparse point cloud clusters within the blind zone range and the newly formed point cloud clusters at the boundary of the blind zone range; determine the sparse point cloud positions in each frame based on each sparse point cloud cluster, and determine the newly formed target positions in each frame based on each newly formed point cloud cluster, where each newly formed point cloud cluster in each frame is the point cloud cluster that appears at the boundary of the blind zone range in the frame; based on the historical tracking data of the disappeared target object and the map data of the target intersection, predict the initial positions of the disappeared target object in each frame within the target time range based on the target time; based on the sparse point cloud positions, newly formed target positions, and initial positions in each frame, determine the target positions of the disappeared target object in each frame.

[0044] In this embodiment, after identifying the disappearing target object in the disappearing target identification stage, the TNT trajectory prediction model enters the disappearing target position prediction stage. In the disappearing target position prediction stage, based on the historical tracking data such as the position, heading angle, and velocity of the disappearing target object and the point cloud blind zone range of the lidar, the target time when the disappearing target object reappears within the point cloud perception range of the lidar is determined. Then, from the point cloud clusters of each frame time range uploaded by the lidar with the target time as the reference, sparse point cloud clusters within the point cloud blind zone range and newly formed point cloud clusters at the boundary of the point cloud blind zone range are obtained. Based on each frame time... The system identifies sparse point cloud clusters and newly generated point cloud clusters, determining the positions of each sparse point cloud and each newly generated target at each frame. Based on the predicted initial position of the disappearing target object at each frame within the subsequent target time range, the system incorporates the newly generated target positions and sparse point cloud positions identified at each frame within the subsequent target time range as constraints. The newly generated target positions are used as references for predicting the target position and endpoint position of the disappearing target object's trajectory. This method determines the target positions of the disappearing target object at each frame, thereby reducing computational complexity and ensuring the accuracy of target tracking and trajectory prediction, as well as the continuity of identification. Specifically, when determining the target positions of the disappearing target object at each frame based on the sparse point cloud positions, newly generated target positions, and initial positions at each frame, the following methods can be used, but are not limited to:

[0045] For each frame, from the initial positions corresponding to that frame, select initial positions whose distance to at least one sparse point cloud position or newly generated target position is no greater than a distance threshold as candidate positions for that frame. From the newly generated target positions corresponding to that frame, select newly generated target positions whose distance to at least one initial position is no greater than a distance threshold as first unknown positions for that frame. From the sparse point cloud positions corresponding to that frame, select sparse point cloud positions whose distance to at least one initial position is no greater than a distance threshold as second unknown positions for that frame. Based on the candidate positions, first unknown positions, and second unknown positions corresponding to that frame, determine the target positions of the disappeared target object at that frame.

[0046] It is worth mentioning that if no new target position with a distance of no more than a distance threshold is found from the new target positions corresponding to that frame time, then it is determined that none of the new target positions corresponding to that frame time belong to the disappeared target objects. The target time is then delayed to obtain the new target time. Furthermore, from the point cloud clusters collected by the lidar in each frame time range based on the new target time, sparse point cloud clusters within the point cloud blind zone and new point cloud clusters at the boundary of the point cloud blind zone are continuously acquired to screen for new target positions until a new target position with a distance of no more than a distance threshold is found from the point cloud clusters in a frame time.

[0047] Step 103: Based on the target positions of the disappearing target object at each frame time, reconstruct the trajectory of the disappearing target object to obtain the target motion trajectory of the disappearing target object.

[0048] In this embodiment, the TNT trajectory prediction model predicts the target positions of the disappearing target object at each frame time in the disappearing target position prediction stage, and then enters the disappearing target trajectory prediction stage. In the disappearing target trajectory prediction stage, based on the target positions of the disappearing target object at each frame time, the trajectory of the disappearing target object is reconstructed to obtain the target motion trajectory of the disappearing target object. Specifically, it can adopt, but is not limited to, the following methods:

[0049] First, multiple initial motion trajectories of the disappearing target object are obtained by trajectory prediction based on the target positions of the disappearing target object at each frame time. Specifically, after generating candidate motion trajectories using a single-peak distribution method based on the target positions of the disappearing target object at each frame time, multiple initial motion trajectories of the disappearing target object are determined based on the position distribution of each candidate motion trajectory and the target positions of the disappearing target object at each frame time.

[0050] Then, the maximum entropy scoring method is used to score each initial motion trajectory, and the initial motion trajectory with the highest score is determined as the target motion trajectory of the disappearing target object.

[0051] Furthermore, in this embodiment, after reconstructing the trajectory of the disappeared target object based on the target positions of the disappeared target object at each frame time to obtain the target motion trajectory of the disappeared target object, the newly emerging target object corresponding to the newly emerging target position in the target motion trajectory can be identified as the disappeared target object, and the identity of the disappeared target object can be assigned to the newly emerging target object corresponding to the newly emerging target position in the target motion trajectory to continue target tracking.

[0052] In addition, when the identity of the disappearing target object disappears, the last tracking data of the Kalman tracker of the disappearing target object is recorded. Based on the target positions on the target motion trajectory of the disappearing target object, the last tracking data of the Kalman tracker of the disappearing target object is iteratively updated. The updated Kalman tracker is then assigned to the new target object corresponding to the new target position in the target motion trajectory to continue target tracking.

[0053] The following provides a detailed description of the smart intersection target tracking method provided in this application embodiment. In this smart intersection target tracking method, the TNT trajectory prediction model is used to identify disappearing target objects. Based on the initial position of the disappearing target object at each frame within the subsequent target time range, constraints are added using the newly identified target position and sparse point cloud position at each frame within the subsequent target time range. The newly identified target position is used as a reference for predicting the target position and endpoint position of the disappearing target object's trajectory. This determines the target positions of the disappearing target object at each frame within the subsequent target time range. Based on these target positions, trajectory reconstruction is performed to obtain the target trajectory of the disappearing target object, thereby reducing computational complexity and ensuring the accuracy of target tracking and trajectory prediction, as well as the continuity of identity identification. In this application embodiment, the TNT trajectory prediction model mainly includes three stages: disappearing target identification stage, disappearing target position prediction stage, and disappearing target trajectory reconstruction stage. Wherein:

[0054] I. Disappearing Target Identification Stage:

[0055] Since traffic targets follow road signs in intersection perception, map data plays a crucial role in target tracking and trajectory prediction. Map data determines travel time and influences the positional relationships between traffic targets. Therefore, in target tracking and trajectory prediction, position points on the lane centerlines are uniformly sampled from the map data of the target intersection and used as the initial position of the target. Assuming the target does not leave the lane (e.g., the outer lane of a two-lane road), if the target is successfully tracked in the previous frame, in the next frame, the target may enter the blind zone of the lidar point cloud or a sparse area with poor point cloud quality, causing the target's identification to be interrupted or changed, resulting in the target disappearing. At this point, the target's possible trajectory is one of four: going straight, changing lanes to the left, decelerating, or braking suddenly. Based on this, the target's position and trajectory are predicted and recorded in real time using lane line data from the map data and historical tracking data of the disappeared target (e.g., speed, position, heading angle).

[0056] In this embodiment of the application, high-precision map data of the target intersection is first collected, including the fixed point cloud blind area of ​​the intersection such as lane lines, zebra crossings, and traffic signs. The edge of the effective point cloud perception range of the lidar is set as the target generation position. That is, when the target object enters the point cloud perception range of the lidar, it will be tracked as a new target.

[0057] If the identification of a target object is interrupted or changed within the point cloud perception range of the lidar or near the point cloud blind zone, the target object is considered a disappeared target object, and its historical tracking data, such as position, heading angle, and velocity, is recorded. Similarly, if a new target appears within the point cloud perception area of ​​the lidar or near the point cloud blind zone, the new target should be re-matched with the identification of the disappeared target object, rather than being assigned a new identification. This avoids unreasonable tracking caused by the disappearance or appearance of identification in the dark.

[0058] II. Prediction of the location of the vanishing target:

[0059] For ease of calculation, in this embodiment of the application, the target coordinate system is transformed from the Cartesian coordinate system to the Frenet coordinate system. The position tracking and trajectory prediction of the target object perceived and identified by the lidar are performed. That is, the sparse point cloud position, the position of the newly formed target, the initial position of the disappearing target object and the target position mentioned later all belong to the position data under the Frenet coordinate system.

[0060] Furthermore, in this embodiment, an improvement is made on the traditional TNT trajectory prediction model, and a TNT trajectory prediction model suitable for tracking disappearing targets is proposed. First, VectorNet is used to encode high-precision map data and target objects to obtain the global features of the target object to be predicted, which are then used for subsequent decoding, thereby completing target tracking and trajectory prediction. The traditional TNT trajectory prediction model includes three stages: (1) Target prediction: Multiple initial positions that the target object may reach are continuously generated on the lane line based on the current position of the target object, and then the most likely target position is selected from each initial position; (2) Target motion estimation: The trajectory distribution of the target object is estimated; (3) Scoring and selection: The multiple predicted motion trajectories are sorted, and a set of target motion trajectories with the highest likelihood score is selected.

[0061] Compared with traditional TNT trajectory prediction models, the TNT trajectory prediction model in this application focuses on selecting the correct target location from various possible locations such as the location of the newly formed target and the location of the sparse point cloud to match the disappearing target object, so as to ensure the continuity of the target movement trajectory of the disappearing target object and ensure the accuracy of subsequent traffic data analysis. Based on this, the traditional TNT trajectory prediction model is improved in this embodiment. The TNT trajectory prediction model provided in this embodiment includes three stages: disappearing target identification stage, disappearing target location prediction stage, and disappearing target trajectory prediction stage. In the disappearing target identification stage, when tracking target objects at the target intersection within the point cloud perception range of the LiDAR based on the point cloud clusters of the target intersection collected by the LiDAR in the current frame, if it is determined that a target object disappearing due to interruption or change of identity identification is present within the point cloud perception range of the LiDAR, then the target object with interrupted or changed identity identification is identified as the disappearing target object. In the disappearing target location prediction stage, based on the historical tracking data such as the position, heading angle, and speed of the disappearing target object and the blind zone range of the LiDAR point cloud, the target time after the disappearing target object reappears within the point cloud perception range of the LiDAR is determined, and the data uploaded from the LiDAR is then used to predict the disappearing target location. In the point cloud clusters within each frame of the target time range, using the target time as a reference, sparse point cloud clusters within the point cloud blind zone and newly formed point cloud clusters at the boundaries of the point cloud blind zone are obtained. Based on each sparse point cloud cluster and each newly formed point cloud cluster in each frame, the positions of each sparse point cloud and each newly formed target are determined. Furthermore, based on predicting the initial position of the disappearing target object in each frame of the subsequent target time range, the positions of newly formed targets and sparse point clouds identified in each frame of the subsequent target time range are added as constraints. The positions of newly formed targets are used as references for predicting the target position and the endpoint position of the disappearing target object's trajectory, thus determining the target position of the disappearing target object in each frame. In the disappearing target trajectory prediction stage, the trajectory of the disappearing target object is reconstructed based on its position in each frame to obtain the target trajectory of the disappearing target object. This eliminates the need for target tracking and trajectory prediction for each initial point, as required by traditional TNT trajectory prediction models, thereby reducing computational complexity while ensuring the accuracy of target tracking and trajectory prediction and the continuity of identity identification.

[0062] See Figure 2 As shown, lane lines are first identified from the map data corresponding to the target intersection, and the possible initial trajectory of the disappearing target object is generated according to the lane line data (e.g., lane center line). Figure 2 The blue dashed line in the image), and the center position of the lane line sampled from the lane line data (e.g., the ...). Figure 2The yellow dot in the image represents the historical tracking position of the disappeared target object across various frames within the historical timeframe. P =[s0,s -1 ,s -2 ,...,s -T [, and based on the historical position s of the disappeared target object at each frame within the historical time range.] P =[s0,s -1 ,s -2 ,...,s -T Predict the initial position s of the disappearing target object at each frame within the future target time range. F =[s1,s2,...,s r Specifically, assume τ(C) P (x, y) represents the initial position (x, y) of the vanishing target object in its direction of movement within the target time range, so multiple initial positions can be generated:

[0063]

[0064] As can be seen from the above formula, traditional TNT estimation and prediction models require oversampling a large number of lane center locations as input, for example, n=1000, to increase the coverage of potential future locations; then, a small number of lane center locations are retained as output, for example, the top 50, for further processing. This results in a large computational load and depends on the accuracy of the lane center locations. To reduce the computational load and improve prediction accuracy, this application proposes an improved target location prediction method, which mainly includes the following:

[0065] 1. Record historical tracking data of the disappeared target object, including position, velocity, acceleration, heading angle, and identification. Based on the historical tracking data of the disappeared target object and the blind zone range of the LiDAR point cloud, preliminarily calculate the target time t when the disappeared target object reappears within the LiDAR point cloud perception range. Obtain point cloud clusters for each frame within the target time range with target time t as the reference. For example, assuming the frame time when the disappeared target object disappears is 0, the target time range can be 1-t.

[0066] 2. When the disappearing target object disappears into the blind zone of the point cloud, sparse point cloud clusters are detected from the point cloud clusters at each frame within the target time range based on the target time t. If they exist, the location range of the sparse point cloud clusters is obtained through clustering and other methods, and the center position of the sparse point cloud clusters is calculated as the sparse point cloud position.

[0067] 3. Based on historical tracking data such as the position, velocity, acceleration, and heading angle of the disappeared target object, as well as the blind zone range of the lidar point cloud, calculate the target time t for the disappearing target object to reappear. If no new target appears near the blind zone range within time t, it can be determined that the disappeared target object is still within the point cloud sensing range. Dynamically manage the target time t, continuously adjusting it to ensure accurate acquisition of the new target's time. For example, delay the target time t until a new target appears between time t and time t, then update the target time t = t ~. This avoids the problem of the disappeared target object staying in the blind zone range for too long and being lost during tracking, ensuring that the disappeared target object can better match the new target.

[0068] 4. If there are newly formed point cloud clusters or sparse point cloud clusters near the blind zone of the point cloud at one or more times t' before reaching the target time t, then determine the unknown location of the disappeared target object based on the newly formed point cloud clusters or sparse point cloud clusters. The unknown locations include the first unknown location and the second unknown location mentioned above.

[0069] 5. Utilize sparse point cloud clusters within the point cloud blind zone and newly formed point cloud clusters near the point cloud blind zone to filter the initial positions of the disappearing target object. Then, using the initial positions of the disappearing target object, select the sparse point cloud position and the newly formed target position that are most likely to match the disappearing target object.

[0070] 6. The endpoint of the target trajectory of a disappearing target object must be the location of a newly formed target to ensure the continuity of the disappearing target object's identity. Because the location of a newly formed target or sparse point cloud may belong to another disappearing target object, it is necessary to filter all possible initial positions, newly formed target positions, and sparse point cloud positions. On the one hand, initial positions whose distance from the newly formed target position or sparse point cloud position is no greater than a distance threshold d are selected as candidate positions from the various initial positions of the disappearing target object. On the other hand, newly formed target positions and sparse point cloud positions without nearby candidate positions are removed. This increases the accuracy of position tracking and trajectory prediction of the disappearing target object, reduces computational complexity, and thus yields the target position of the disappearing target object at different frame times.

[0071]

[0072] Wherein, if the new target position (x) new ,y new Unable to find a candidate position that meets the distance threshold d {(x n ,y n )+(Δx n ,Δy nIf the newly generated target location is not considered to belong to the disappeared target object, then step 3 continues until a new target location that meets the conditions is obtained.

[0073] The target position of a disappearing target object at different frame times consists of two parts: one part is the newly generated target position and / or sparse point cloud position corresponding to subsequent frame times, and the other part is the candidate position corresponding to subsequent frame times, which is composed of initial positions whose distance from the newly generated target position and / or sparse point cloud position is no greater than a distance threshold d. Because it is impossible to determine whether the newly generated target position and sparse point cloud position belong to the disappearing target object, they are also called unknown positions. Unknown positions include the aforementioned first unknown position and second unknown position. If all unknown positions within the point cloud blind zone at a certain frame time cannot match candidate positions, then all initial positions predicted at that frame time need to be used as candidate positions.

[0074] In the TNT trajectory prediction model, the target position distribution p(τ|(S)) of the vanished target object is predicted for each frame within the future target time range. p C p When performing this process, firstly, probability distribution prediction is performed on the initial positions that satisfy the distance threshold d to obtain the most likely candidate positions. Then, probability distribution prediction is performed on the newly generated target positions and sparse point cloud positions that satisfy the distance threshold d to obtain the most likely unknown positions (including the aforementioned first and second unknown positions). Based on the candidate positions and unknown positions, the target position is determined, specifically including:

[0075] The following formula is used to calculate the probability distribution of the n initial positions that meet the distance threshold d, and the top Z1 initial positions are selected as candidate positions;

[0076]

[0077] The following formula is used to calculate the probability distribution of the m newly generated target locations and sparse point cloud locations that meet the distance threshold d, and the top Z2 newly generated target locations and sparse point cloud locations are taken as unknown locations (including the first and second unknown locations mentioned above);

[0078] p(τ new m |(s p ,c p ))=π(τ new m |(s p ,c p ))

[0079] The historical tracking data S is calculated using the following formula. P Candidate position τ g Unknown position τnew (including the aforementioned first and second unknown locations) and environmental factors c P (e.g., signs, traffic lights, etc.) are merged as feature data and then the probability distribution is calculated to obtain the final target location;

[0080] p(τ n |(s p ,c p ))=π(τ n |(s p ,c p ,τ new ,τ g ))

[0081] in, f(), N(), v() are two-layer MLPs.

[0082] Unlike traditional TNT trajectory prediction models, the TNT trajectory prediction model in this embodiment predicts two parts for the target position: one part is the candidate position that is closest to the movement position of the disappearing target object, and the other part is the unknown position (including the aforementioned first unknown position and second unknown position) that is closest to the movement position of the disappearing target object, calculated based on the position of the newly formed target and the position of the sparse point cloud. These target positions may be one or multiple, so the generated initial trajectory may also be multiple, which requires further screening.

[0083] III. Reconstruction of the Disappearing Target Trajectory:

[0084] The previous stage of predicting the location of the disappearing target predicted the locations of each target object, but did not provide a complete motion trajectory. For example, a selected target location may be highly likely to belong to the disappearing target object, but the complete motion trajectory to this target location may be unreasonable or discontinuous, so further filtering is required.

[0085] Assume the interaction between the disappearing target object and the surrounding environment of the target intersection (e.g., other target objects, lane information, lane it is in, etc.), i.e., the environmental factors are c. P =[c0,c -1 ,c -2 ,..,c T Then, the probability distribution of the candidate motion trajectory obtained after trajectory prediction is expressed as p(s). F |(s P ,c P In this process, trajectory prediction may yield multiple motion trajectories. For example, a vehicle approaching an intersection may turn into the far right or middle lane when turning left.

[0086]

[0087] Where N(x) is a two-layer MLP, p(S) F |τ,(S p C p The candidate motion trajectories are generated by using a single-peak distribution method based on the target positions of the disappeared target object at each frame within the future target time range.

[0088] Then, based on the probability distribution of candidate motion trajectories, multiple initial motion trajectories of the disappeared target object are determined, and the target motion trajectory is filtered from these initial motion trajectories. First, the features of each initial motion trajectory are calculated, similar to dense target encoding, and then passed to a decoder of a 2-layer MLP. Maximum entropy scoring is used to score the M initial motion trajectories, where g(x) is a two-layer MLP.

[0089]

[0090] The initial trajectory with the highest final score is taken as the target trajectory. The tracked target trajectory is output. If the target object disappears and the disappearance location is unreasonable, i.e., not at the edge of the LiDAR point cloud perception range, the full trajectory is obtained by matching the last tracking data at the time of disappearance with the predicted target trajectory. The identity of the disappeared target object is assigned to the new target object corresponding to the new target position on the target trajectory to achieve continuous tracking.

[0091] Although newly generated target objects have accurate identification, the traditional TNT trajectory prediction model still needs to initialize the Kalman tracker for new target objects to perform new tracking because the Kalman tracker stops updating when the target object disappears. This not only affects accuracy but also requires several iterations of the initialized Kalman tracker in subsequent frames to achieve accurate tracking. To optimize this issue, this application proposes a method for continuing to track newly generated target objects using the Kalman tracker of the disappeared target object.

[0092] In the above method, the identity identifiers of newly created target objects and disappearing target objects are matched, and the motion trajectory of disappearing target objects is predicted and reconstructed. Therefore, when a disappearing target object disappears, its Kalman tracker is recorded, and the target position on the reconstructed target motion trajectory is processed and regarded as the motion position of subsequent disappearing target objects. The Kalman tracker is continuously updated using these positions until the position of the newly created target is updated, and the updated Kalman tracker is assigned to the newly created target object for target tracking.

[0093] Assuming the disappearance time of the target object is t1, record the last state of its Kalman tracker (i.e., the last tracking data);

[0094] x t+1 =Fx t +Bu t +W t

[0095] Where F is the state transition equation, which transforms the state at time t to the state at time t+1, and obtains the target position corresponding to the target's trajectory at the next time t+1, denoted as Z. t+1 Based on the error p of the optimal estimate at time k-1 t-1 Calculate the error p of the predicted value at time k t Update the covariance matrix P of the state vector;

[0096]

[0097]

[0098]

[0099] Since the trajectory of a single target has already been predicted, matching and tracking are unnecessary. The only requirement is to reduce the estimated values ​​and candidate points at different times and minimize errors to ensure the accuracy of the Kalman tracker. When updating to the location of a newly formed target, its observations are no longer candidate points but rather the results of subsequent identification. The Kalman tracker is then assigned to the newly formed target object for continued tracking, completing the global tracking of the disappearing target object.

[0100] In summary, please refer to Figure 3 As shown, high-precision map data of the target intersection is pre-collected, including lane lines, zebra crossings, traffic signs, etc., and the effective point cloud perception range and fixed point cloud blind zone range of the LiDAR at the target intersection are pre-measured. During target tracking, the point cloud clusters of the target intersection reported by the LiDAR are acquired in real time, and the latest frame is used for target identification. The identification algorithm can be a target identification algorithm such as PVRCNN or PointPillar. After the target object is identified, a Kalman tracker is initialized according to the number of target objects to track the target objects and assign a unique identity to each target object. The tracking data such as the identity, position, speed, and heading angle of the target object are stored. When the identity of a target object is interrupted, it is determined whether it is at the edge of the point cloud perception range. If it is, it is considered to have disappeared normally; if not, it is determined to be a disappeared target object that failed to be tracked, and the historical tracking data such as the position, speed, and heading angle of the disappeared target object are recorded.

[0101] The system records the frame containing the vanished target object and then screens subsequent frames one by one. If a new or sparse point cloud cluster exists near the vanished target object in a subsequent frame at time 'a', the center positions of the new and sparse point cloud clusters are recorded. Based on historical tracking data such as the lane, position, speed, and time 'a' of the vanished target object, position prediction is performed to generate multiple possible initial positions for the vanished target object at time 'a'. These initial positions are then matched with the positions of the new and sparse point clouds. A distance threshold 'd' is set; if the initial position is not greater than the distance threshold 'd', the match is considered successful. If no match is found, the search for new and sparse point cloud positions continues in subsequent frames.

[0102] The target position is obtained by filtering the successfully matched initial position, newly generated target position, and sparse point cloud position. Trajectory prediction is performed on the target position to generate one or more initial motion trajectories. These initial motion trajectories include those generated based on the newly generated target position and sparse point cloud position, as well as those generated based on the initial position. Because the positions of the newly generated point cloud cluster and sparse point cloud cluster are sparser than the initial position, these initial motion trajectories may contain unreasonable or incomplete phenomena. Maximum entropy scoring is performed on these initial motion trajectories to estimate and select the most reasonable target motion trajectory.

[0103] When a disappearing target object vanishes, its Kalman tracker is recorded. The target position on the reconstructed target trajectory is then processed and treated as the position of a subsequent disappearing target object. These target positions are used to continuously update the Kalman tracker until a new target position is reached. The updated Kalman tracker is then assigned to the new target object. Finally, the identity of the disappearing target object is assigned to the new target object, and the normal motion trajectory determined during the tracking of the new target object and the predicted target motion trajectory during the prediction of the disappearing target object are stitched together to complete the global tracking of the disappearing target object.

[0104] In this way, by using sparse point cloud clusters and newly formed point cloud clusters as reference points for location trajectory prediction, the target motion trajectory of a disappearing target object can be predicted, which can improve the continuity and accuracy of target perception data at smart intersections while reducing the amount of computation.

[0105] Based on the above embodiments, this application provides a smart intersection target tracking device, see below. Figure 4 As shown, the smart intersection target tracking device 400 provided in this application embodiment includes at least:

[0106] The disappearing target identification module 401 is used to track the target object at the target intersection within the point cloud perception range of the laser radar based on the point cloud cluster of the target intersection at the current frame time. If it is determined that a target object with a disappearing identity mark appears within the point cloud perception range of the laser radar, then the target object with the disappearing identity mark is identified as the disappearing target object.

[0107] The disappearing target location prediction module 402 is used to determine the target time when the disappeared target object reappears within the point cloud perception range of the LiDAR, based on the historical tracking data before the disappearance of the disappearing target object's identity and the blind zone range of the LiDAR point cloud; from the point cloud clusters collected by the LiDAR in each frame of the target time range based on the target time, it obtains sparse point cloud clusters within the blind zone range and newly formed point cloud clusters at the boundary of the blind zone range; it determines the position of each sparse point cloud cluster in each frame based on each sparse point cloud cluster, and determines the position of each newly formed target in each frame based on each newly formed point cloud cluster; wherein, each newly formed point cloud cluster in each frame is the point cloud cluster that appears at the boundary of the blind zone range in the frame; based on the historical tracking data of the disappeared target object and the map data of the target intersection, it predicts the initial position of the disappeared target object in each frame of the target time range based on the target time; based on the sparse point cloud positions, newly formed target positions, and initial positions in each frame, it determines the target positions of the disappeared target object in each frame.

[0108] The vanishing target trajectory reconstruction module 403 is used to reconstruct the trajectory of the vanishing target object based on the target positions of the vanishing target object at each frame time to obtain the target motion trajectory of the vanishing target object.

[0109] In one possible implementation, the disappearing target location prediction module 402 is specifically configured to, for each frame time, select initial positions from the initial positions corresponding to the frame time whose distance from at least one sparse point cloud position or newly generated target position is no greater than a distance threshold as candidate positions corresponding to the frame time, select newly generated target positions from the newly generated target positions corresponding to the frame time whose distance from at least one initial position is no greater than a distance threshold as first unknown positions corresponding to the frame time, and select sparse point cloud positions from the sparse point cloud positions corresponding to the frame time whose distance from at least one initial position is no greater than a distance threshold as second unknown positions corresponding to the frame time, and determine the various target positions of the disappearing target object at each frame time based on the candidate positions, the first unknown positions, and the second unknown positions corresponding to the frame time.

[0110] In one possible implementation, the disappearing target location prediction module 402 is specifically used to determine that none of the newly emerging target locations corresponding to the frame time belong to the disappearing target object if no newly emerging target location with a distance not greater than a distance threshold is found from the newly emerging target locations corresponding to the frame time, and to obtain a new target time by delaying the target time. It also continuously acquires sparse point cloud clusters within the point cloud blind zone and newly emerging point cloud clusters at the boundary of the point cloud blind zone from the point cloud clusters collected by the lidar within the new target time range based on the new target time, and performs new target location screening until a newly emerging target location with a distance not greater than a distance threshold is found from the point cloud clusters of a frame time.

[0111] In one possible implementation, the disappearing target trajectory reconstruction module 403 is specifically used to predict the trajectory of the disappearing target object based on each target position of the disappearing target object at each frame time to obtain multiple initial motion trajectories of the disappearing target object.

[0112] The maximum entropy scoring method is used to score each initial motion trajectory, and the initial motion trajectory with the highest score is determined as the target motion trajectory of the disappearing target object.

[0113] In one possible implementation, the disappearing target trajectory reconstruction module 403 is specifically used to generate each candidate motion trajectory based on each target position of the disappearing target object at each frame time using a single-peak distribution method; and to determine multiple initial motion trajectories of the disappearing target object based on each candidate motion trajectory and the position distribution of each target position of the disappearing target object at each frame time.

[0114] In one possible implementation, the smart intersection target tracking device 400 provided in this application embodiment further includes:

[0115] The identity identification continuation module 404 is used to identify the newly emerging target object corresponding to the newly emerging target position in the target movement trajectory as the disappearing target object, and to assign the identity identification of the disappearing target object to the newly emerging target object corresponding to the newly emerging target position in the target movement trajectory to continue target tracking.

[0116] In one possible implementation, the smart intersection target tracking device 400 provided in this application embodiment further includes:

[0117] The tracker continuation module 405 is used to record the last tracking data of the Kalman tracker of the disappeared target object when the identity of the disappeared target object disappears; based on the target positions on the target motion trajectory of the disappeared target object, iteratively update the last tracking data of the Kalman tracker of the disappeared target object; and assign the updated Kalman tracker to the new target object corresponding to the new target position in the target motion trajectory to continue target tracking.

[0118] It should be noted that the principle of the intelligent intersection target tracking device 400 provided in this application embodiment to solve the technical problem is similar to the intelligent intersection target tracking method provided in this application embodiment. Therefore, the implementation of the intelligent intersection target tracking device 400 provided in this application embodiment can refer to the implementation of the intelligent intersection target tracking method provided in this application embodiment, and the repeated parts will not be described again.

[0119] After introducing the smart intersection target tracking method and device provided in the embodiments of this application, the electronic device provided in the embodiments of this application will be briefly introduced next.

[0120] The electronic devices provided in this application embodiment may be, but are not limited to, computers, servers, etc. For specific details, please refer to [link / reference]. Figure 5 As shown, the electronic device 500 provided in this application embodiment includes at least a processor 501, a memory 502, and a computer program stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program, it implements the above-mentioned smart intersection target tracking method provided in this application embodiment.

[0121] The electronic device 500 provided in this application embodiment may further include a bus 503 connecting different components (including processor 501 and memory 502). The bus 503 represents one or more types of bus structures, including memory bus, peripheral bus, local area bus, etc.

[0122] Memory 502 may include readable media in the form of volatile memory, such as random access memory (RAM) 5021 and / or cache memory 5022, and may further include read-only memory (ROM) 5023. Memory 502 may also include a program tool 5025 having a set (at least one) of program modules 5024, including but not limited to an operating subsystem, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0123] Processor 501 can be a single processing element or a collective term for multiple processing elements. For example, processor 501 can be a microcontroller unit (MCU), a central processing unit (CPU), or one or more integrated circuits configured to implement the intelligent intersection target tracking method provided in the embodiments of this application. Specifically, processor 501 can be a general-purpose processor, including but not limited to CPUs, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0124] Electronic device 500 can communicate with one or more external devices 504 (e.g., keyboard, remote control, etc.), and also with one or more devices that enable a user to interact with electronic device 500 (e.g., mobile phone, computer, etc.), and / or with devices that enable electronic device 500 to communicate with one or more other electronic devices (e.g., router, modem, etc.). This communication can be performed through input / output (I / O) interface 505. Electronic device 500 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) through network adapter 506. Figure 5 As shown, network adapter 506 communicates with other modules of electronic device 500 via bus 503. It should be understood that, although... Figure 5 As not shown, other hardware and / or software modules may be used in conjunction with electronic device 500, including but not limited to microcode, device drivers, redundant processors, external disk drive arrays, Redundant Arrays of Independent Disks (RA) subsystems, tape drives, and data backup storage subsystems.

[0125] It should be noted that, Figure 5 The electronic device 500 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0126] Furthermore, this application embodiment also provides a computer-readable storage medium storing computer instructions. When executed by a processor, these computer instructions implement the aforementioned smart intersection target tracking method provided in this application embodiment. Specifically, the computer instructions can be built into or installed in the processor, so that the processor can implement the aforementioned smart intersection target tracking method provided in this application embodiment by executing the built-in or installed computer instructions.

[0127] In addition, the smart intersection target tracking method provided in the embodiments of this application can also be implemented as a program product, which includes program code. When the program code is executed by a processor, it implements the smart intersection target tracking method provided in the embodiments of this application.

[0128] The program product provided in this application embodiment can be any combination of one or more readable media, wherein the readable media can be a readable signal medium or a readable storage medium, and the readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. Specifically, more specific examples of readable storage media (a non-exhaustive list) include electrical connections with one or more wires, portable disks, hard disks, RAM, ROM, erasable programmable read-only memory (EPROM), optical fibers, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0129] The program product provided in this application embodiment can be a CD-ROM and include program code, and can also run on electronic devices such as computers and servers. However, the program product provided in this application embodiment is not limited to this. In this application embodiment, the readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, device, or apparatus.

[0130] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0131] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0132] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0133] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A smart intersection target tracking method, characterized in that, include: Based on the point cloud cluster of the target intersection collected by the lidar at the current frame moment, when tracking the target object of the target intersection within the point cloud perception range of the lidar, if it is determined that a target object with a missing identity tag appears within the point cloud perception range of the lidar, then the target object with a missing identity tag is determined as the missing target object. Based on the historical tracking data of the disappeared target object before its identity disappeared and the blind zone range of the lidar point cloud, the target time when the disappeared target object reappears within the point cloud perception range of the lidar is determined. From the point cloud clusters collected by the lidar at each frame within the target time range based on the target time, sparse point cloud clusters within the point cloud blind zone and newly formed point cloud clusters at the boundary of the point cloud blind zone are obtained. Based on each of the sparse point cloud clusters at each frame, the positions of each sparse point cloud at each frame are determined, and the positions of each newly formed target at each frame are determined based on each of the newly formed point cloud clusters at each frame. Wherein, each of the newly formed point cloud clusters at each frame is a point cloud cluster that appears at the boundary of the point cloud blind zone at that frame. Based on the historical tracking data of the disappeared target object and the map data of the target intersection, predict the initial position of the disappeared target object at each frame within the target time range based on the target time. Based on the sparse point cloud positions, newly generated target positions, and initial positions at each frame time, the target positions of the disappearing target objects at each frame time are determined. Based on the target positions of the vanished target object at each frame time, the trajectory of the vanished target object is reconstructed to obtain the target motion trajectory of the vanished target object.

2. The intelligent intersection target tracking method as described in claim 1, characterized in that, Based on the sparse point cloud positions, newly generated target positions, and initial positions at each frame time, the target positions of the disappearing target objects at each frame time are determined, including: For each frame time, from the initial positions corresponding to the frame time, the initial positions whose distance to at least one sparse point cloud position or the newly generated target position is not greater than a distance threshold are selected as candidate positions corresponding to the frame time. From the newly generated target positions corresponding to the frame time, the newly generated target positions whose distance to at least one initial position is not greater than the distance threshold are selected as first unknown positions corresponding to the frame time. From the sparse point cloud positions corresponding to the frame time, the sparse point cloud positions whose distance to at least one initial position is not greater than the distance threshold are selected as second unknown positions corresponding to the frame time. Based on the candidate positions, the first unknown positions, and the second unknown positions corresponding to the frame time, the target positions of the disappeared target object at each frame time are determined.

3. The intelligent intersection target tracking method as described in claim 2, characterized in that, Also includes: If, from the newly generated target locations corresponding to the frame time, no newly generated target location is found whose distance to at least one of the initial locations is not greater than the distance threshold, then it is determined that none of the newly generated target locations corresponding to the frame time belong to the disappeared target object. The target time is then delayed to obtain a new target time. Furthermore, from the point cloud clusters collected by the lidar within the new target time range based on the new target time, sparse point cloud clusters within the point cloud blind zone and newly generated point cloud clusters at the boundary of the point cloud blind zone are continuously acquired to filter the newly generated target locations until a newly generated target location whose distance to at least one of the initial locations is not greater than the distance threshold is found in the point cloud clusters of a frame time.

4. The intelligent intersection target tracking method as described in claim 1, characterized in that, Based on the target positions of the vanished target object at each frame time, trajectory reconstruction is performed on the vanished target object to obtain the target motion trajectory of the vanished target object, including: Based on the trajectory prediction of the disappearing target object at each target position at each frame time, multiple initial motion trajectories of the disappearing target object are obtained; Each initial motion trajectory is scored using the maximum entropy scoring method, and the initial motion trajectory with the highest score is determined as the target motion trajectory of the disappearing target object.

5. The intelligent intersection target tracking method as described in claim 4, characterized in that, Based on trajectory prediction of the vanishing target object at each target position in each frame, multiple initial motion trajectories of the vanishing target object are obtained, including: Based on the target positions of the disappeared target object at each frame time, each candidate motion trajectory is generated using a single-peak distribution method; Based on each of the candidate motion trajectories and the positional distribution of the disappearing target object at each of the target positions at each frame time, multiple initial motion trajectories of the disappearing target object are determined.

6. The intelligent intersection target tracking method as described in any one of claims 1-5, characterized in that, Also includes: The newly emerging target object corresponding to the newly emerging target position in the target motion trajectory is identified as the disappeared target object, and the identity of the disappeared target object is assigned to the newly emerging target object corresponding to the newly emerging target position in the target motion trajectory to continue target tracking.

7. The intelligent intersection target tracking method as described in any one of claims 1-5, characterized in that, Also includes: Record the last tracking data of the Kalman tracker of the disappeared target object when the identity of the disappeared target object disappears; Based on the target positions on the target motion trajectory of the disappeared target object, the last tracking data of the Kalman tracker of the disappeared target object is iteratively updated; The updated Kalman tracker is assigned to the new target object corresponding to the new target position in the target motion trajectory to continue target tracking.

8. A smart intersection target tracking device, characterized in that, include: The disappearing target identification module is used to track the target object at the target intersection within the point cloud perception range of the LiDAR based on the point cloud cluster of the target intersection at the current frame time. If it is determined that a target object with a disappearing identity tag appears within the point cloud perception range of the LiDAR, then the target object with the disappearing identity tag is identified as the disappearing target object. The disappearing target location prediction module is used to determine the target time when the disappearing target object reappears within the point cloud perception range of the lidar, based on the historical tracking data of the disappearing target object before its identity disappeared and the point cloud blind zone range of the lidar. From the point cloud clusters collected by the lidar in each frame within the target time range based on the target time, sparse point cloud clusters within the point cloud blind zone and newly formed point cloud clusters at the boundary of the point cloud blind zone are obtained. Based on each of the sparse point cloud clusters in each frame, the positions of each sparse point cloud in each frame are determined, and the positions of each newly formed target in each frame are determined based on each of the newly formed point cloud clusters in each frame. Each of the newly formed point cloud clusters in each frame is a point cloud cluster that appears at the boundary of the point cloud blind zone in that frame. Based on the historical tracking data of the disappeared target object and the map data of the target intersection, the initial positions of the disappeared target object in each frame within the target time range based on the target time are predicted. Based on the sparse point cloud positions, newly generated target positions, and initial positions at each frame time, the target positions of the disappearing target objects at each frame time are determined. The vanishing target trajectory reconstruction module is used to reconstruct the trajectory of the vanishing target object based on the target positions of the vanishing target object at each frame time to obtain the target motion trajectory of the vanishing target object.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the intelligent intersection target tracking method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the smart intersection target tracking method as described in any one of claims 1-7.

Images