Multi-target tracking method and device in occlusion scene, equipment and storage medium

Abstract

The application provides a multi-target tracking method and device in an occlusion scene, equipment and a storage medium, the method comprising: obtaining point cloud clusters of tracking objects before occlusion; filtering the point cloud clusters of the tracking objects to obtain probability distribution information of predicted positions; processing point cloud data of each tracking object after occlusion to obtain tracking point cloud clusters; and determining target tracking objects corresponding to the tracking point cloud clusters according to Mahalanobis distances between the tracking point cloud clusters and the probability distribution information. The application can better depict possible positions of tracking objects in an occlusion scene by predicting the probability distribution of positions, and therefore can better achieve multi-target tracking in the occlusion scene based on the Mahalanobis distances between the tracking point cloud clusters and the position probability distribution, thereby improving the tracking effect of multi-targets in the occlusion scene.

Description

Technical Field

[0001] This application relates to the field of target tracking technology, specifically to a method, apparatus, device, and storage medium for multi-target tracking in occluded scenarios. Background Technology

[0002] Target tracking refers to the real-time tracking and identification of moving objects in a scene using technologies such as vision and lidar scanning, and it has been widely used in various fields.

[0003] However, in many practical applications, such as obstacle courses for robots, various occlusions in the scene often affect the acquired scene images or point cloud data, thus impacting object tracking performance. Especially in multi-object tracking, when two objects in close proximity are occluded and re-acquired, the accuracy of re-identifying each tracked object using the intersection-union ratio of the object detection boxes is not ideal. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, device, and storage medium for multi-target tracking in occluded scenarios to address the aforementioned technical problems, thereby resolving the issue that the related technologies do not achieve ideal results in tracking multiple targets in occluded scenarios.

[0005] Firstly, this application provides a multi-target tracking method for occluded scenarios, including:

[0006] Obtain the point cloud clusters of each tracked object in the target scene before occlusion;

[0007] The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the predicted position corresponding to each tracked object;

[0008] The point cloud data of each tracked object appearing in the target scene after occlusion is processed to obtain a tracking point cloud cluster;

[0009] Based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted location, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0010] As a feasible embodiment of this application, the step of filtering the point cloud clusters of each tracked object to obtain the probability distribution information of the predicted position corresponding to each tracked object includes:

[0011] For each tracked object, the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object.

[0012] Based on the Gaussian distribution parameters, a covariance matrix is ​​generated to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object.

[0013] As a feasible embodiment of this application, determining the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector based on the first point cloud cluster and the second point cloud cluster of the tracked object includes:

[0014] The motion state parameters of the tracked object in the preset spatial coordinate system are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object.

[0015] The motion state parameters are statistically analyzed to determine the Gaussian distribution parameters of the tracked object on the first and second axis vectors in a preset spatial coordinate system.

[0016] As a feasible embodiment of this application, the step of generating a covariance matrix based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object to obtain the probability distribution information of the predicted position corresponding to the tracked object includes:

[0017] The covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, thereby obtaining the first probability distribution information of the predicted position of the tracked object at the first prediction time.

[0018] The covariance matrix is ​​updated based on the first probability distribution information to obtain the updated covariance matrix;

[0019] Based on the updated covariance matrix, Kalman filtering is performed on the point cloud clusters of the tracked object to obtain the second probability distribution information of the predicted position of the tracked object at the second prediction time, until the probability distribution information of each predicted position of the tracked object when it appears in the target scene after occlusion is determined.

[0020] As a feasible embodiment of this application, determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes:

[0021] The Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position corresponding to each tracking object is compared.

[0022] If the minimum Mahalanobis distance in the Mahalanobis distance is less than a preset distance threshold, the tracking object corresponding to the minimum Mahalanobis distance is determined as the target tracking object corresponding to the tracking point cloud cluster.

[0023] If the minimum Mahalanobis distance in the Mahalanobis distance is greater than a preset distance threshold, the tracking point cloud cluster is identified as a newly emerging tracking object.

[0024] As a feasible embodiment of this application, determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes:

[0025] The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the motion state corresponding to each tracked object;

[0026] Determine the motion state information corresponding to each of the aforementioned tracking point cloud clusters;

[0027] Based on the first Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the tracking object, and the second Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0028] As a feasible embodiment of this application, determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes:

[0029] Determine the occlusion duration for each of the tracked objects;

[0030] The Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position corresponding to the tracking object is weighted according to the weight corresponding to the occlusion duration to obtain the weighted Mahalanobis distance.

[0031] The target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects based on the weighted Mahalanobis distance.

[0032] Secondly, this application provides a multi-target tracking device for occluded scenarios, comprising:

[0033] The acquisition module is used to acquire the image of the document to be identified and the corresponding portrait background image of the document;

[0034] The processing module is used to process the document image and the portrait background image respectively using a trained feature extraction model to obtain the first image feature of the document image and the second image feature of the portrait background image;

[0035] The determination module is used to determine the multi-target tracking result of the document image in an occluded scenario based on the feature similarity between the first image features and the second image features.

[0036] Thirdly, this application also provides a computer device, the computer device comprising:

[0037] One or more processors;

[0038] Memory; and

[0039] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the multi-target tracking method in occlusion scenarios provided above.

[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to execute the multi-target tracking method in the occlusion scenario provided above.

[0041] Fifthly, embodiments of this application provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the multi-target tracking method in an occlusion scenario described above.

[0042] The multi-target tracking method in occlusion scenarios provided in this application's embodiments filters the point cloud clusters of each tracked object in the target scene before occlusion to determine the probability distribution information describing the possible positions of the tracked objects at each predicted location. This facilitates subsequent clustering to obtain the tracking point cloud clusters in the target scene after occlusion. The method can then use the Mahalanobis distance between the tracking point cloud clusters and the probability distribution information of the predicted positions of each tracked object to determine the target tracked object corresponding to each tracking point cloud cluster. This is equivalent to existing technologies relying on object detection boxes and matching the tracked object based on the intersection-union ratio (IoU) of the object's detection box and the detection box at the predicted position. Since the position probability distribution better characterizes the possible positions of the tracked object in occlusion scenarios compared to the detection box at the predicted position, multi-target tracking based on the Mahalanobis distance between the tracking point cloud clusters and the position probability distribution performs better in occlusion scenarios, thus improving the tracking effect of multi-target tracking in occlusion scenarios. Attached Figure Description

[0043] 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.

[0044] Figure 1 This is a flowchart illustrating the steps of the multi-target tracking method in an occlusion scenario provided in an embodiment of this application.

[0045] Figure 2 A flowchart illustrating the steps for predicting the probability distribution of the location of a tracked object, as provided in an embodiment of this application;

[0046] Figure 3 This is a schematic diagram illustrating the modeling effect of object motion changes in the polar coordinate space coordinate system provided in the embodiments of this application;

[0047] Figure 4 This is a schematic flowchart illustrating the steps of performing Kalman filtering on dynamically updated covariance matrix pairs according to an embodiment of this application.

[0048] Figure 5 This is a flowchart illustrating the steps for determining the target tracking object corresponding to a tracking point cloud cluster based on Mahalanobis distance, as provided in an embodiment of this application.

[0049] Figure 6 This application provides a schematic flowchart illustrating the steps for achieving association prediction based on the state dimension in an embodiment of the present application.

[0050] Figure 7 This application provides a flowchart illustrating a step for determining a target tracking object based on Mahalanobis distance weighting at occlusion time, as provided in an embodiment of the present application.

[0051] Figure 8 This is a schematic diagram of the structure of a multi-target tracking device in an occlusion scenario provided in an embodiment of this application;

[0052] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly and completely described 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 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.

[0054] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0055] In the description of this application, the word "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0056] To facilitate understanding of the multi-target tracking method in occluded scenarios provided in this application embodiment, the relevant application background of the multi-target tracking method in occluded scenarios is first explained. Specifically, target tracking refers to collecting data of the target scene through technologies such as vision or LiDAR, determining object information in the scene based on image recognition or point cloud clustering, and then tracking the continuously collected object information in the scene in real time through state prediction, such as through the state transition equation in Kalman filtering. It is currently widely used in fields such as intelligent monitoring and autonomous driving. However, in some practical application scenarios with occlusion, such as various obstacle course competition scenarios for robots, occlusion can cause the loss of some object information in the collected target scene data, which will affect the tracking effect to a certain extent. Especially in the process of multi-target tracking, when two objects in close positions are occluded and re-collected and identified, the recognition accuracy of determining each tracked object by the intersection-union ratio of the object detection box and the predicted position in related technologies is not ideal.

[0057] To address the tracking failure issues that easily occur in occluded scenarios, this application introduces distribution parameters related to object dynamics during state prediction to describe the probability distribution of each possible predicted position of the tracked object after occlusion. It then uses the Mahalanobis distance between the point cloud clusters where the tracked object reappears in the target scene and the probability distribution of the tracked object at each possible predicted position to measure the correlation between the point cloud clusters and the tracked object. This allows for accurate identification of the tracked object corresponding to each point cloud cluster, ensuring effective target tracking in occluded scenarios. The multi-target tracking method in occluded scenarios typically refers to a computer program embedded in a multi-target tracking device. This device typically refers to a processor embedded in a computer device. The multi-target tracking device in the computer device executes the computer program corresponding to the multi-target tracking method provided in this application embodiment to achieve multi-target tracking in occluded scenarios.

[0058] For details, please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of the multi-target tracking method in an occlusion scenario provided in the embodiments of this application, including steps S110 to S140:

[0059] S110, Obtain the point cloud clusters of each tracked object in the target scene before occlusion.

[0060] Before the tracked object in the target scene is occluded, similar to conventional target tracking methods, data of the target scene can be collected using vision or LiDAR to track the object when it is not occluded. For example, an image acquisition device, such as a camera, can be used to acquire images of the target scene, and target detection and recognition can be performed on the images. Then, the tracking of each object is completed based on the intersection over union (IoU) of the detection boxes between objects in images acquired at adjacent time points. Of course, considering that this application is mainly applied to multi-target tracking in occluded scenes, especially for tracking and matching the tracked object before and after occlusion, and that visual information is more susceptible to occlusion interference than LiDAR, in order to further improve the recognition effect of multi-target tracking in occluded scenes, this embodiment of the application chooses to use LiDAR to acquire the point cloud clusters of the tracked object before occlusion.

[0061] Specifically, acquiring point cloud clusters of each tracked object in a target scene before occlusion based on LiDAR typically involves the following steps:

[0062] (1) Initialize the point cloud data acquisition process, including the estimation of the background point cloud in the target scene. Then, ICP (Iterative Closest Point) registration is performed by manually calibrating feature points to obtain the approximate position of each lidar. Then, the pose is accurately registered based on the point-to-plane algorithm to obtain the precise pose of the lidar.

[0063] (2) Using a 3D voxel map, background point cloud filtering is achieved by differentiating the background based on the occupancy of the voxel, resulting in clean point cloud data. Specifically, this involves creating a background voxel map for each lidar, then using the background point cloud to occupy the voxel map to obtain a coarser filtering map. Based on the lidar's pose, a detection ray emitted from the lidar itself is constructed to detect whether each voxel can be directly perceived. All imperceptible voxels are occupied in the voxel map, effectively filtering out scattering noise points from the lidar. Then, the occupied voxel portions are expanded. The above steps effectively improve the filtering efficiency of the background point cloud. Due to the use of a voxel map, after initialization, the time complexity of a single-point query is O(1), and the space complexity is only linearly related to the site size and voxel density.

[0064] (3) By selecting a suitable point cloud clustering algorithm, the filtered point cloud is clustered to locate all the tracking objects from the point cloud data of the target scene. The clustering algorithm can be set according to actual needs. For example, in occluded scenes, it is not required to predetermine the number of clusters and a certain amount of noise points can be tolerated. Therefore, as a common and feasible implementation scheme, a density-based clustering method with noise (DBSCAN) can be selected to cluster the filtered point cloud data. Of course, in the actual process, considering that the point cloud density decreases with the increase of the LiDAR distance, it is difficult to determine the clustering radius and the minimum number of clustered points of the DBSCAN algorithm. Therefore, as a further optional scheme of this application, during the clustering of point cloud data by DBSCAN, the relevant parameters of DBSCAN can be dynamically updated based on the inverse quadratic relationship between the LiDAR point cloud density and the distance to ensure the clustering effect of the point cloud, thereby obtaining the point cloud clusters of each tracking object in the target scene before occlusion.

[0065] It's important to note that because LiDAR collects and clusters point clouds of the target scene in real time, it can obtain point cloud clusters at multiple different times. Before the tracked object is occluded, by predicting the state of these tracked objects and then using the intersection-union ratio (IUU) of the object detection boxes detected at adjacent times, each tracked object can be effectively identified and identified with a corresponding ID (identification information) for subsequent matching when occlusion occurs. Therefore, the point cloud clusters of each tracked object before occlusion typically include point cloud cluster data from multiple times before the tracked object is occluded.

[0066] S120, the point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the predicted position corresponding to each tracked object.

[0067] In this embodiment, since the point cloud data of the tracked object cannot be observed after it is occluded, it is impossible to use the intersection-over-union (IoU) ratio of the object detection box and the previous object detection box to achieve tracking and matching of the tracked object. Therefore, it is usually necessary to consider state prediction of the point cloud clusters of the tracked object to obtain the possible predicted position of the tracked object, thereby achieving position matching of the tracked object before and after occlusion. For example, as a common and feasible implementation, the point cloud clusters of the tracked object can be filtered to determine the state transition equation. Based on the point cloud cluster data of the tracked object before occlusion, the predicted position of the tracked object at the time of occlusion can be updated and calculated, thereby completing the tracking and matching of the point cloud clusters of the tracked object after occlusion.

[0068] However, in the above process, conventional waves are only used to predict the predicted position of the tracked object. In the technical solution of this application, the motion change model of the tracked object is reasonably constructed to predict the probability distribution information of the predicted position of the tracked object, that is, it describes the probability distribution of the tracked object appearing in each different position.

[0069] Specifically, to facilitate understanding of the above content and to make the probability distribution of the predicted positions more consistent with the motion state of the tracked object, this application embodiment will further provide an implementation scheme for predicting the probability distribution of the tracked object's position by combining specific motion change modeling. For details, please refer to... Figure 2 , Figure 2 The following is a detailed flowchart illustrating the steps for predicting the probability distribution of the location of a tracked object, as provided in an embodiment of this application.

[0070] The solution provided in this application requires performing similar modeling and processing for each tracked object detected before occlusion. Therefore, for the sake of simplicity, this application only uses one tracked object as an example for illustration. For other tracked objects, processing can be performed based on similar steps, and this application will not repeat the details. Specifically, it includes steps S210 to S220:

[0071] S210, determine the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector based on the first point cloud cluster and the second point cloud cluster of the tracked object.

[0072] In this embodiment of the application, as can be seen from the foregoing description, since tracking an object before occlusion requires point cloud cluster data at multiple times, after each tracked object is determined before occlusion, point cloud clusters of the tracked object at multiple different times can be collected simultaneously. Therefore, the point cloud clusters of the tracked object at adjacent times can be used as the first point cloud cluster and the second point cloud cluster to determine the distribution parameters of the tracked object on different axis vectors in a specified space, such as the radial axis vector and angular axis vector in the polar coordinate system, or the horizontal axis vector and vertical axis vector in the Cartesian coordinate system.

[0073] Of course, in order to make the prediction results fit the motion state of the tracked object in space, the Gaussian distribution parameters of the tracked object on the first and second axis vectors are typically determined by using the first and second point cloud clusters of the tracked object to determine its motion state. Based on this motion state information, Gaussian distribution parameters in the corresponding space are generated to facilitate further prediction of a more accurate position probability distribution. Specifically, determining the Gaussian distribution parameters of the tracked object on the first and second axis vectors based on the first and second point cloud clusters includes:

[0074] The motion state parameters of the tracked object in the preset spatial coordinate system are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object.

[0075] The motion state parameters are statistically analyzed to determine the Gaussian distribution parameters of the tracked object on the first and second axis vectors in a preset spatial coordinate system.

[0076] Specifically, to facilitate understanding of the above content, this application will use a polar coordinate space coordinate system as an example for illustration. For details, please refer to... Figure 3 , Figure 3 This is a schematic diagram illustrating the modeling effect of object motion changes in the polar coordinate system provided in this embodiment of the application. Details are as follows.

[0077] Please see Figure 3 xk-1 x represents the position information of the tracked object observed at the time preceding time k. k This corresponds to the position information of the tracked object observed at time k, hat x k This corresponds to applying the state transition equation to x. k-1 The predicted position information of the tracked object at time k is processed. Clearly, Δx (positive in the velocity direction) and θ (positive in the counter-clockwise direction) can reflect the deviation in the prediction result caused by changes in the motion state of the tracked object. If we assume that all probability distributions in the model follow a normal distribution, then we can approximately consider that both Δx and θ satisfy a normal distribution with an expectation of 0, and that they are independent of each other, i.e., Δx ~ N(0, σ1). 2 ) and θ~N(0,σ2) 2 To obtain the optimal estimation results, σ1 and σ2 should be estimated by statistically analyzing Δx and θ of the tracked object in the real situation. That is, σ1 and σ2 are estimated using the statistical results of the first and second point cloud clusters at multiple adjacent time points. Specifically, the limiting cases of Δx and θ can be analyzed using the 3σ principle for estimation, where the formula for calculating Δx is:

[0078]

[0079] a k It is acceleration, therefore, it can be obtained through the maximum acceleration a max The estimated value is σ1, that is:

[0080]

[0081] Similarly, for σ2, the maximum angular acceleration ω is also passed through... max It is estimated that:

[0082]

[0083] Based on the above method, the bivariate normal distribution of the predicted position in polar coordinates can be obtained. Of course, besides obtaining the bivariate normal distribution of the predicted position in radial axis vector Δx and angular axis vector θ in polar coordinate space, the corresponding bivariate normal distribution of the predicted position can also be obtained in Cartesian coordinate space. Specifically, this can be further approximated by considering the predicted velocity direction of the tracked object... Establish a right-handed Cartesian coordinate system with the positive x-axis as the axis. Then, when θ is sufficiently small, the x′ component of the predicted value can be considered as... y′ axis component They are independent, meaning that the expected value of the predicted value can be determined in this coordinate system. And since polar coordinate space and the aforementioned Cartesian coordinate space have the following relationship formula:

[0084]

[0085] in, Therefore, combining the aforementioned bivariate normal distribution of the predicted position in polar coordinates, we can obtain:

[0086]

[0087] Since the above is based on the predicted velocity direction of the tracked object. A right-handed rectangular coordinate system x′Oy′ is established with the positive x-axis as the reference direction. Therefore, it can be further transformed into the xOy spatial coordinate system of the target scene using the spatial coordinate system transformation formula, thereby obtaining the covariance matrix Q of the predicted position. k :

[0088]

[0089] in, It is the basis transformation matrix from the x′Oy′ coordinate system to the xOy coordinate system, and is usually related to the predicted velocity direction and predicted position of the tracked object.

[0090] S220, a covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object.

[0091] Based on the foregoing descriptions, after generating the Gaussian distribution parameters that describe the predicted position of the tracked object, the covariance matrix Q of the predicted position can be generated from these parameters. k The covariance matrix Q k It can be used for Kalman filtering of point cloud clusters of tracked objects to predict the corresponding predicted position of the tracked object in real time. Specifically, the update formula of Kalman filtering is as follows:

[0092]

[0093] in, and That is, the expected value P describes the position of the tracked object at time k and time k-1. k and P k-1 That is, the covariance matrix corresponding to the position of the tracked object at time k and time k-1, describing the distribution of the tracked object's position, while F k These are the kinematic parameters of a target that obeys rigid body kinematics and will move at a uniform linear velocity without being subjected to any external forces. Specifically:

[0094]

[0095] It can be seen that the covariance matrix Q, determined through the aforementioned method, k In such cases, even when observation data about the target is obstructed or cannot be obtained, the future position of the target can be predicted based on the existing state using Kalman filtering. In this situation, the object's inertia is utilized, and it begins to drift inertially according to the estimated velocity result from the previous moment, meaning that the covariance matrix will also increase directionally.

[0096] As can be seen, in the aforementioned process, the modeling model used in the embodiments of this application is similar to the common fixed Q model. k The usage of Kalman filtering differs from that of Q. k The probability distribution will change due to the velocity at time k, thus allowing the estimated probability distribution to better reflect reality. Furthermore, since the parameter estimation model is based on a real-world scenario, the estimated probability distribution has practical significance in fitting the actual motion of the tracked object. This makes it comparable based on Mahalanobis distance during subsequent tracking and matching. In other words, during the process of generating the covariance matrix based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object to obtain the probability distribution information of the predicted position of the tracked object, the covariance matrix will be dynamically updated in real time. For details, please refer to [link to relevant documentation]. Figure 4 , Figure 4 This is a flowchart illustrating the steps of performing Kalman filtering on dynamically updated covariance matrix pairs according to an embodiment of this application. Specifically, it includes steps S410 to S430:

[0097] S410, a covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, thereby obtaining the first probability distribution information of the predicted position of the tracked object at the first prediction time.

[0098] In this embodiment, due to the duration and location of the occlusion of the tracked object, when performing Kalman filtering on the point cloud cluster of the tracked object, Kalman filtering is first performed based on the covariance matrix generated by the Gaussian distribution parameters to obtain the first probability distribution information of the predicted position of the tracked object at the first prediction time, which is the next prediction time.

[0099] S420, the covariance matrix is ​​updated according to the first probability distribution information to obtain the updated covariance matrix.

[0100] Unlike conventional unoccluded scenarios where the covariance matrix can be updated based on the observation position information, in this embodiment, since the observation position information of the tracked object cannot be obtained when occluded, it is necessary to update the covariance matrix based on the probability distribution information of the predicted position to obtain the updated covariance matrix. Since the sum of two normal distributions still follows a normal distribution, in this embodiment, the covariance matrix describing the probability distribution of the predicted position is updated based on the probability distribution information. The resulting updated covariance matrix still describes the probability distribution of the predicted position at the next time step using a normal distribution.

[0101] S430, Kalman filtering is performed on the point cloud cluster of the tracked object based on the updated covariance matrix to obtain the second probability distribution information of the predicted position of the tracked object at the second prediction time, until the probability distribution information of each predicted position of the tracked object when it appears in the target scene after occlusion is determined.

[0102] In this embodiment, by combining the updated covariance matrix with further Kalman filtering of the point cloud cluster coordinates of the tracked object, the second probability distribution information for determining the predicted position of the tracked object at the next prediction time after occlusion (i.e., the second prediction time) can be updated and used to further update the covariance matrix Q. k The algorithm is dynamically updated until the tracked object reappears in the target scene after being occluded. At this point, the probability distribution information describing the possible location of the tracked object is determined for subsequent tracking and matching.

[0103] Of course, in the specific Kalman filtering model, in addition to the aforementioned Q... k In addition, it usually includes the covariance matrix R of the observation error. k In contrast, since the observation error is relatively fixed and usually originates from upper-level observation sources, it is assumed that the observation error will not change with the target's motion state. Therefore, the limit of the observation error can be reasonably estimated using the 3σ principle as the target radius r, that is:

[0104]

[0105] Furthermore, other well-known common knowledge in the Kalman filtering process will not be elaborated upon in the embodiments of this application.

[0106] S130, the point cloud data of each tracked object appearing in the target scene after occlusion is processed to obtain a tracking point cloud cluster.

[0107] In this embodiment, when each tracked object reappears in the target scene, that is, when the point cloud data in the target scene can be clustered using the aforementioned DSBCAN clustering method to generate several new tracking point cloud clusters, it is necessary to redetermine the tracking object corresponding to each tracking point cloud cluster, or to determine that the point cloud cluster is a new tracking object. Especially in multi-target tracking scenarios, it is necessary to determine the association mapping relationship of each tracking object before and after occlusion. The method of processing and clustering the point cloud data of the tracking object appearing in the target scene after occlusion to obtain the tracking point cloud clusters is similar to the method of determining the point cloud clusters of the tracking object before occlusion, and will not be repeated here.

[0108] S140, based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position, determine the target tracking object corresponding to the tracking point cloud cluster from the tracking objects.

[0109] In this embodiment of the application, as can be seen from the foregoing description, since the mathematical modeling process is based on the real scene and the estimation of each parameter is also based on the motion state of the tracked object, the estimated probability distribution has the practical significance of fitting the real motion of the tracked object. Therefore, the target tracked object corresponding to each tracking point cloud cluster can be determined from the tracked object based on the similarity between the actual detected tracking point cloud cluster after occlusion, that is, the observed value and the obtained predicted position, that is, the predicted value.

[0110] It should be noted that due to occlusion, the collected point cloud clusters are discontinuous. Therefore, it is impossible to determine the target tracking object corresponding to each tracking point cloud cluster by calculating the intersection-over-union ratio (IoU) between the tracking point cloud cluster and the predicted position using conventional IoU methods. Furthermore, in the technical solution of this application embodiment, the predicted position exists in the form of a probability distribution information following a normal distribution. That is, it is impossible to use conventional methods such as Euclidean distance to calculate the similarity between the tracking point cloud cluster and the predicted position. Therefore, this application embodiment uses Mahalanobis distance, which describes the distance between a point and a distribution, to achieve matching between the tracking point cloud cluster and the tracking object. Specifically, taking spatial location as an example, the probability distribution information of a tracking point cloud cluster position x0 and the predicted position of a tracking object... In this case, the formula for calculating the Mahalanobis distance between the two is as follows:

[0111]

[0112] Mahalanobis distance reflects the probability density of matching between the current object and historical objects. Comparing Mahalanobis distances essentially involves maximum likelihood estimation. Qualitatively, due to its Euclidean distance characteristics (i.e., on the extension of the expected value), the farther the distance from the expected value, the larger the Mahalanobis distance. Under equal Euclidean distance, the closer the motion state is to the velocity direction, the larger the Mahalanobis distance. Therefore, by analyzing the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information at different predicted positions, the target tracking object corresponding to the tracking point cloud cluster can be determined. This is especially true when there are multiple tracking objects in similar positions before occlusion. Due to the superposition of multiple expected normal distributions, there will be sufficient differences in the probability distribution information at the predicted positions. In this case, the Mahalanobis distance can more accurately determine the tracking point cloud cluster of each tracking object after occlusion. Specifically, the comparison based on Mahalanobis distance can usually determine the target tracking object by determining the minimum Mahalanobis distance between the tracking point cloud cluster and the probability distribution information at different predicted positions. For details, please refer to [link to relevant documentation]. Figure 5 , Figure 5 This application provides a flowchart illustrating the steps for determining the target tracking object corresponding to a tracking point cloud cluster based on Mahalanobis distance, specifically including steps S510 to S530:

[0113] S510, compare the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position corresponding to each tracking object.

[0114] In this embodiment of the application, for each tracking point cloud cluster, the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position corresponding to each tracking object can be calculated based on the Mahalanobis distance calculation formula provided above. Then, these calculated Mahalanobis distances are compared to determine the minimum Mahalanobis distance.

[0115] S520, if the minimum Mahalanobis distance in the Mahalanobis distance is less than a preset distance threshold, the tracking object corresponding to the minimum Mahalanobis distance is determined as the target tracking object corresponding to the tracking point cloud cluster.

[0116] In this embodiment of the application, when the minimum Mahalanobis distance is less than a preset distance threshold, that is, when it is considered that the probability distribution of the predicted position of a certain tracking point cloud cluster and a certain tracking object is sufficiently similar, it can be considered that the tracking point cloud cluster and the tracking object are sufficiently similar, that is, the tracking object corresponding to the minimum Mahalanobis distance can be determined as the target tracking object corresponding to the tracking point cloud cluster.

[0117] S530, if the minimum Mahalanobis distance in the Mahalanobis distance is greater than a preset distance threshold, the tracking point cloud cluster is identified as a newly emerging tracking object.

[0118] In this embodiment of the application, corresponding to the foregoing, if the minimum Mahalanobis distance is greater than a certain preset distance threshold, that is, if the similarity between a certain tracking point cloud cluster and a certain most likely tracking object still cannot meet the requirements, then the tracking point cloud cluster can be considered as a newly emerging tracking target. If necessary, other observation information can be combined in subsequent processes, such as combining visual observation information for further matching and registration. This embodiment of the application will not be elaborated here.

[0119] Specifically, the above solution can be described by the following formula:

[0120]

[0121] Among them, T k-1 Let the target being tracked at time k-1 be denoted as .

[0122] And X k This represents the cluster at time k, and this step requires specifying the initial state, which can be considered as...

[0123] Of course, in addition to the aforementioned method of measuring the correlation between tracking point cloud clusters and tracking objects based solely on the Mahalanobis distance between the probability distribution information of the predicted positions of the tracking point cloud clusters and the corresponding tracking objects in the position dimension, as a further feasible implementation of this application, the Mahalanobis distance between tracking point cloud clusters and tracking objects in the motion state dimension can also be considered to more accurately achieve the matching between tracking point cloud clusters and tracking objects. For details, please refer to [link to relevant documentation]. Figure 6 , Figure 6 This application provides a flowchart illustrating steps for achieving association prediction based on the state dimension, specifically including steps S610 to S630:

[0124] S610, the point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the motion state corresponding to each tracked object.

[0125] In this embodiment of the application, as can be seen from the foregoing description, when performing Kalman filtering on the point cloud cluster of the tracked object to determine the probability distribution information of the possible predicted position of each tracked object, when considering the relationship between displacement, velocity and acceleration, and performing reasonable kinematic modeling of the tracked object, the probability distribution information of the motion state corresponding to each tracked object can also be predicted, such as the velocity direction, velocity magnitude, acceleration, etc., which follow a normal distribution.

[0126] S620, determine the motion state information corresponding to each of the tracking point cloud clusters.

[0127] In this embodiment of the application, a certain tracking point cloud cluster can only describe the position information at a certain moment. Therefore, under normal circumstances, further clustering of the point cloud data under the occluded target scene is performed to obtain tracking point cloud clusters at multiple moments. The motion state information corresponding to each tracking point cloud cluster can be determined, such as the aforementioned velocity direction, velocity magnitude and acceleration information.

[0128] S630, based on the first Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the tracking object, and the second Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0129] Compared to directly using the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and different tracking objects at the predicted position in the location dimension to determine the association between the tracking object and the tracking point cloud cluster, this embodiment of the application further combines the Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the predicted tracking object to determine the association between the tracking object and the tracking point cloud cluster more accurately. For example, due to the normalization in the Mahalanobis distance, that is, the distance can be compared under different parameter distributions, the first Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the tracking object, and the second Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the predicted position of the tracking object can be appropriately weighted to obtain the weighted Mahalanobis distance. This weighted distance allows for a more accurate determination of the target tracking object corresponding to the tracking point cloud cluster from the tracking objects. The method for determining the first Mahalanobis distance between the motion state information and the probability distribution information of the motion state corresponding to the tracking object can refer to the aforementioned calculation formula for the Mahalanobis distance between the predicted positions. Alternatively, if the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted positions of multiple tracking objects is small enough and close enough, for example, if the difference between the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted positions of the first and second tracking objects is less than a preset threshold, then analyzing only the position may still lead to the risk of matching failure. Therefore, we can continue to consider the relationship between the Mahalanobis distance between the motion state of the tracking point cloud cluster and the probability distribution information of the corresponding predicted motion states of the first and second tracking objects to determine the target tracking object corresponding to the tracking point cloud cluster.

[0130] By combining the probability distribution information in the motion state dimension with the position dimension mentioned above to achieve the matching of tracking point cloud clusters and tracking objects, the matching accuracy of tracking objects can be further improved, which is especially suitable for multi-target tracking processes in occluded scenarios.

[0131] In addition to the aforementioned implementation scheme that further improves the matching accuracy of the tracked object based on the probability distribution information of the motion state dimension, considering that the probability distribution information of the predicted position is obtained by superimposing the probability distributions of the predicted positions at multiple times, that is, as the occlusion time increases, the error of the position prediction will accumulate. That is, under the same Euclidean distance, the tracked object will lose tracking, that is, the longer the occlusion time, the larger the Mahalanobis distance can be considered. Therefore, as a further optional implementation scheme of this application, when determining the target tracked object corresponding to the tracked point cloud cluster from the tracked objects based on the Mahalanobis distance between the probability distribution information of the tracked point cloud cluster and the predicted positions of different tracked objects, considering that different tracked objects are occluded at different times, that is, the tracked objects may not be occluded at the same time, the occlusion time of the tracked object can be further considered to appropriately optimize the weight of the Mahalanobis distance, thereby further improving the matching accuracy of the tracked object. For details, please refer to Figure 7 , Figure 7 This application provides a flowchart illustrating a step for determining a target tracking object based on Mahalanobis distance weighting at occlusion time, specifically including steps S710 to S720:

[0132] S710, determine the occlusion duration of each tracked object.

[0133] In this embodiment, the occlusion duration of the tracked object can be understood as the occlusion duration between the loss of a cluster of a tracked object and the appearance of a new tracked point cloud cluster in the target scene during the clustering process of point cloud data of the target scene. It can usually be determined based on a timer or based on the period of point cloud data acquisition.

[0134] S720, the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position corresponding to the tracking object is weighted according to the weight corresponding to the occlusion duration, so as to obtain the weighted Mahalanobis distance.

[0135] In this embodiment of the application, as can be seen from the foregoing description, since the probability distribution of the predicted position is accumulated, that is, as the occlusion duration increases, the error of the predicted position corresponding to the tracked object will also accumulate. Therefore, the Mahalanobis distance between the probability distribution information of the tracked point cloud cluster and the predicted position corresponding to the tracked object can be weighted based on the weight corresponding to the occlusion duration. The longer the occlusion duration, the higher the corresponding weight coefficient can be set. That is, the occlusion duration can be positively correlated with the weight, so that the weighted Mahalanobis distance between the probability distribution information of the tracked point cloud cluster and the predicted position corresponding to the tracked object is larger.

[0136] S730, determine the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the weighted Mahalanobis distance.

[0137] In this embodiment, the Mahalanobis distance between the probability distribution information of the predicted position of the tracking point cloud cluster and the tracking object is weighted by incorporating the weights corresponding to the occlusion duration. This weighted Mahalanobis distance is then used to determine the target tracking object corresponding to the tracking point cloud cluster from among the tracking objects, thus more accurately identifying the target tracking object. For example, when the Mahalanobis distances between the tracking point cloud cluster and the predicted position probability distribution information of two tracking objects, such as the third and fourth tracking objects, are relatively close, if the occlusion duration of the third tracking object is longer than that of the fourth tracking object, it is assumed that the cumulative error between the predicted position probability distribution information of the fourth tracking object and the actual position of the fourth tracking object is lower than that of the third tracking object. Therefore, the tracking point cloud cluster is considered more likely to be the fourth tracking object. In this case, the weighted Mahalanobis distance based on the occlusion duration is smaller, leading to the selection of the fourth tracking object as the target tracking object.

[0138] Currently, the aforementioned weighting of Mahalanobis distance based on occlusion duration can also be applied to the weighting of Mahalanobis distance between motion state information. That is, the features of the above embodiments can be appropriately combined based on actual needs. The implementation scheme after combination will not be described in detail in this application embodiment.

[0139] The multi-target tracking method in occlusion scenarios provided in this application's embodiments filters the point cloud clusters of each tracked object in the target scene before occlusion to determine the probability distribution information describing the possible positions of the tracked objects at each predicted location. This facilitates subsequent clustering to obtain the tracking point cloud clusters in the target scene after occlusion. The method can then use the Mahalanobis distance between the tracking point cloud clusters and the probability distribution information of the predicted positions of each tracked object to determine the target tracked object corresponding to each tracking point cloud cluster. This is equivalent to existing technologies relying on object detection boxes and matching the tracked object based on the intersection-union ratio (IoU) of the object's detection box and the detection box at the predicted position. Since the position probability distribution better characterizes the possible positions of the tracked object in occlusion scenarios compared to the detection box at the predicted position, multi-target tracking based on the Mahalanobis distance between the tracking point cloud clusters and the position probability distribution performs better in occlusion scenarios, thus improving the tracking effect of multi-target tracking in occlusion scenarios.

[0140] To better implement the multi-target tracking method in occlusion scenarios provided in the embodiments of this application, this application also provides a multi-target tracking device in occlusion scenarios, based on the multi-target tracking method in occlusion scenarios provided in the embodiments of this application. Figure 8 As shown, the multi-target tracking device 800 in occluded scenes includes:

[0141] The acquisition module 810 is used to acquire the point cloud clusters of each tracked object in the target scene before occlusion.

[0142] The prediction module 820 is used to filter the point cloud clusters of each of the tracked objects to obtain the probability distribution information of the predicted position corresponding to each tracked object.

[0143] Processing module 830 is used to process the point cloud data of each of the tracked objects appearing in the target scene after occlusion, to obtain a tracking point cloud cluster;

[0144] The determination module 840 is used to determine the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information.

[0145] In some embodiments of this application, the prediction module 820 is used to determine the Gaussian distribution parameters of the tracking object on the first axis vector and the second axis vector based on the first point cloud cluster and the second point cloud cluster of the tracking object for each tracking object.

[0146] Based on the Gaussian distribution parameters, a covariance matrix is ​​generated to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object.

[0147] In some embodiments of this application, the prediction module 820 is used to determine the motion state parameters of the tracked object in a preset spatial coordinate system based on the first point cloud cluster and the second point cloud cluster of the tracked object;

[0148] The motion state parameters are statistically analyzed to determine the Gaussian distribution parameters of the tracked object on the first and second axis vectors in a preset spatial coordinate system.

[0149] In some embodiments of this application, the prediction module 820 is used to generate a covariance matrix based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, so as to obtain the first probability distribution information of the predicted position of the tracked object at the first prediction time.

[0150] The covariance matrix is ​​updated based on the first probability distribution information to obtain the updated covariance matrix;

[0151] Based on the updated covariance matrix, Kalman filtering is performed on the point cloud clusters of the tracked object to obtain the second probability distribution information of the predicted position of the tracked object at the second prediction time, until the probability distribution information of each predicted position of the tracked object when it appears in the target scene after occlusion is determined.

[0152] In some embodiments of this application, the determining module 840 is used to compare the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted positions corresponding to each tracking object;

[0153] If the minimum Mahalanobis distance in the Mahalanobis distance is less than a preset distance threshold, the tracking object corresponding to the minimum Mahalanobis distance is determined as the target tracking object corresponding to the tracking point cloud cluster.

[0154] If the minimum Mahalanobis distance in the Mahalanobis distance is greater than a preset distance threshold, the tracking point cloud cluster is identified as a newly emerging tracking object.

[0155] In some embodiments of this application, the determining module 840 is used to perform filtering processing on the point cloud clusters of each tracked object to obtain the probability distribution information of the motion state corresponding to each tracked object;

[0156] Determine the motion state information corresponding to each of the aforementioned tracking point cloud clusters;

[0157] Based on the first Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the tracking object, and the second Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0158] In some embodiments of this application, the determining module 840 is used to determine the occlusion duration of each of the tracked objects;

[0159] The Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position corresponding to the tracking object is weighted according to the weight corresponding to the occlusion duration to obtain the weighted Mahalanobis distance.

[0160] The target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects based on the weighted Mahalanobis distance.

[0161] Specific limitations regarding multi-target tracking devices in occluded scenarios can be found in the limitations of multi-target tracking methods in occluded scenarios described above, and will not be repeated here. Each module in the aforementioned multi-target tracking device for occluded scenarios can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0162] In some embodiments of this application, the multi-target tracking device 800 in occlusion scenarios can be implemented as a computer program, which can be implemented in, for example... Figure 9 The computer device shown is running the program. The computer device's memory can store the various program modules that make up the multi-target tracking device 800 in this occlusion scenario, for example, Figure 8 The diagram shows an acquisition module 810, a prediction module 820, a processing module 830, and a determination module 840. The computer program comprised of these modules causes the processor to execute the steps in the multi-target tracking method for occlusion scenarios described in the various embodiments of this application.

[0163] For example, Figure 9 The computer equipment shown can be used as follows Figure 8The acquisition module 810 of the multi-target tracking device 800 in the illustrated occlusion scenario executes step S110. The computer device can execute step S120 via the prediction module 820. The computer device can execute step S130 via the processing module 830. The computer device can execute step S140 via the determination module 840. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computational and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with external computer devices via a network connection. When the computer program is executed by the processor, it implements a multi-target tracking method in an occlusion scenario.

[0164] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0165] In some embodiments of this application, a computer device is provided, including one or more processors; memory; and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the processor to perform the following steps:

[0166] Obtain the point cloud clusters of each tracked object in the target scene before occlusion;

[0167] The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the predicted position corresponding to each tracked object;

[0168] The point cloud data of each tracked object appearing in the target scene after occlusion is processed to obtain a tracking point cloud cluster;

[0169] Based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted location, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0170] In some embodiments of this application, a computer-readable storage medium is provided, storing a computer program that is loaded by a processor, causing the processor to perform the following steps:

[0171] Obtain the point cloud clusters of each tracked object in the target scene before occlusion;

[0172] The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the predicted position corresponding to each tracked object;

[0173] The point cloud data of each tracked object appearing in the target scene after occlusion is processed to obtain a tracking point cloud cluster;

[0174] Based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted location, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

[0175] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0176] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0177] The foregoing has provided a detailed description of a multi-target tracking method, apparatus, computer device, and storage medium in an occlusion scenario provided by the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-target tracking method in occluded scenarios, characterized in that, include: Obtain the point cloud clusters of each tracked object in the target scene before occlusion; The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the predicted position corresponding to each tracked object; The point cloud data of each tracked object appearing in the target scene after occlusion is processed to obtain a tracking point cloud cluster; Based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted location, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects; The step of filtering the point cloud clusters of each tracked object to obtain the probability distribution information of the predicted position corresponding to each tracked object includes: For each tracked object, the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object. The covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object; Determining the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector based on the first point cloud cluster and the second point cloud cluster of the tracked object includes: The motion state parameters of the tracked object in the preset spatial coordinate system are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object. The motion state parameters are statistically analyzed to determine the Gaussian distribution parameters of the tracked object on the first and second axis vectors in a preset spatial coordinate system. The step of generating a covariance matrix based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object, includes: The covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, thereby obtaining the first probability distribution information of the predicted position of the tracked object at the first prediction time. The covariance matrix is ​​updated based on the first probability distribution information to obtain the updated covariance matrix; Based on the updated covariance matrix, Kalman filtering is performed on the point cloud clusters of the tracked object to obtain the second probability distribution information of the predicted position of the tracked object at the second prediction time, until the probability distribution information of each predicted position of the tracked object when it appears in the target scene after occlusion is determined.

2. The method according to claim 1, characterized in that, The step of determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes: The Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position corresponding to each tracking object is compared. If the minimum Mahalanobis distance in the Mahalanobis distance is less than a preset distance threshold, the tracking object corresponding to the minimum Mahalanobis distance is determined as the target tracking object corresponding to the tracking point cloud cluster. If the minimum Mahalanobis distance in the Mahalanobis distance is greater than a preset distance threshold, the tracking point cloud cluster is identified as a newly emerging tracking object.

3. The method according to claim 1, characterized in that, The step of determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes: The point cloud clusters of each tracked object are filtered to obtain the probability distribution information of the motion state corresponding to each tracked object; Determine the motion state information corresponding to each of the aforementioned tracking point cloud clusters; Based on the first Mahalanobis distance between the motion state information corresponding to the tracking point cloud cluster and the probability distribution information of the motion state corresponding to the tracking object, and the second Mahalanobis distance between the tracking point cloud cluster and the probability distribution information of the predicted position, the target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects.

4. The method according to any one of claims 1 to 3, characterized in that, The step of determining the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position includes: Determine the occlusion duration for each of the tracked objects; The Mahalanobis distance between the probability distribution information of the tracking point cloud cluster and the predicted position corresponding to the tracking object is weighted according to the weight corresponding to the occlusion duration to obtain the weighted Mahalanobis distance. The target tracking object corresponding to the tracking point cloud cluster is determined from the tracking objects based on the weighted Mahalanobis distance.

5. A multi-target tracking device for occluded scenarios, characterized in that, include: The acquisition module is used to acquire point cloud clusters of each tracked object in the target scene before occlusion; The prediction module is used to filter the point cloud clusters of each tracked object to obtain the probability distribution information of the predicted position corresponding to each tracked object. The processing module is used to process the point cloud data of each tracked object appearing in the target scene after occlusion, to obtain a tracking point cloud cluster; The determination module is used to determine the target tracking object corresponding to the tracking point cloud cluster from the tracking objects based on the Mahalanobis distance between the tracking point cloud cluster and the probability distribution information; The step of filtering the point cloud clusters of each tracked object to obtain the probability distribution information of the predicted position corresponding to each tracked object includes: For each tracked object, the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object. The covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object; Determining the Gaussian distribution parameters of the tracked object on the first axis vector and the second axis vector based on the first point cloud cluster and the second point cloud cluster of the tracked object includes: The motion state parameters of the tracked object in the preset spatial coordinate system are determined based on the first point cloud cluster and the second point cloud cluster of the tracked object. The motion state parameters are statistically analyzed to determine the Gaussian distribution parameters of the tracked object on the first and second axis vectors in a preset spatial coordinate system. The step of generating a covariance matrix based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud clusters of the tracked object, thereby obtaining the probability distribution information of the predicted position corresponding to the tracked object, includes: The covariance matrix is ​​generated based on the Gaussian distribution parameters to perform Kalman filtering on the point cloud cluster of the tracked object, thereby obtaining the first probability distribution information of the predicted position of the tracked object at the first prediction time. The covariance matrix is ​​updated based on the first probability distribution information to obtain the updated covariance matrix; Based on the updated covariance matrix, Kalman filtering is performed on the point cloud clusters of the tracked object to obtain the second probability distribution information of the predicted position of the tracked object at the second prediction time, until the probability distribution information of each predicted position of the tracked object when it appears in the target scene after occlusion is determined.

6. A computer device, characterized in that, The computer device includes: One or more processors; Memory; and One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the multi-target tracking method in an occlusion scenario as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program, which is loaded by a processor to execute the multi-target tracking method in an occlusion scenario as described in any one of claims 1 to 4.

Images