Automatic driving tracking method and system for long-time shielding scene and storage medium

Abstract

The invention relates to the field of automatic driving tracking, in particular to an automatic driving system tracking method and system for a long-time shielding scene and a storage medium. A mainstream target detector is used to carry out target detection on traffic participants in a scene to obtain a detection frame, a Kalman filter based on a CV motion model is used to carry out trajectory prediction, a virtual trajectory of the traffic participants in a sheltered period is constructed to be used for secondarily updating parameters of the Kalman filter, optimal registration is carried out on IOU of the prediction frame and the detection frame, and a prediction result is obtained. Assigning different tracking IDs to the detection frames corresponding to the traffic participants at the current moment; the problem that when one traffic participant is shielded for a long time, it is difficult to accurately predict the trajectory of the traffic participant through a detection-based tracking algorithm DBT algorithm at the present stage is solved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving tracking, and specifically to an autonomous driving tracking method, system and storage medium for long-time occlusion scenarios. Background Technique

[0002] The multi-object tracking algorithms in the industry are divided into two categories. One is the end-to-end JDE (Joint Detection and Embedding) that combines detection, and the detection-based multi-object tracking DBT (Detection-Based Tracking). JDE is generally a model that combines detection and tracking. Although it reduces the loss of information transmission, it requires a large amount of data for training, and there are many problems in deploying the algorithm model at the edge due to computing power. Most of the industrial community uses DBT. Most of these methods rely on KF (Kalman filter). KF itself assumes that the motion state of an object is linear motion in a very short time, and the prediction noise and observation noise both follow a Gaussian distribution.

[0003] In the prior art, for example, Chinese Patent CN113723190A discloses a multi-object tracking method for synchronous moving targets, but it can only accept traffic participant occlusions within a very short time. When a traffic participant is occluded for a long time, the linear estimation of the Kalman filter will become very inaccurate due to the extension of the traffic participant occlusion time. This is because when there is no measurement value to update the Kalman filter parameters, the Kalman filter will use the prior state estimate as the posterior state update, and the state noise of the traffic participant will continuously accumulate, resulting in error accumulation during the occlusion of the traffic participant. The accumulated error will cause a change in the predicted motion direction of the traffic participant in practical applications. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an autonomous driving tracking method, system and storage medium for long-time occlusion scenarios, which solves the problem that it is difficult for the DBT algorithm to accurately predict the trajectory of a traffic participant when a traffic participant is occluded for a long time.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An autonomous driving tracking method for long-time occlusion scenarios, the tracking method specifically includes the following steps:

[0007] S1. Obtain the detection results of all traffic participants at any moment based on the object detection algorithm, and assign a tracking ID to the detection results at the initial moment; the detection results include the detection box of the traffic participant, the center coordinates of the detection box, the size of the detection box, and the confidence level;

[0008] S2. Predict the prediction boxes of different existing tracking IDs at the current moment using the Kalman filter based on the trajectory data under different existing tracking IDs.

[0009] S3. Calculate the IOU values between the detection box at the current moment and the prediction boxes of different existing tracking IDs in turn, and form an IOU cost matrix. Set the values less than the IOU threshold in the cost matrix to infinity. Use the Hungarian algorithm for the cost matrix to obtain the optimal matching. Assign the tracking ID of the prediction box corresponding to the optimal matching to the detection box corresponding to the optimal matching, and assign a new tracking ID to the detection box whose IOU values are all less than the IOU threshold.

[0010] S4. Determine whether all the tracking IDs except the new tracking ID at the current moment appeared at the previous moment.

[0011] If so, return to step S1.

[0012] If not, mark the trajectory where the detection box corresponding to the tracking ID that did not appear at the previous moment is located as an uncontinuously matched trajectory. This tracking ID does not include the new tracking ID.

[0013] S5. Update the parameter values of the Kalman filter according to the detection boxes of the tracking ID corresponding to the uncontinuously matched trajectory at the current moment and the last appearance to generate a virtual trajectory between the two moments.

[0014] Preferably, in step S1, the object detection algorithm is the CenterNet algorithm.

[0015] Preferably, in step S2, it specifically includes the following steps:

[0016] S21. Define the state quantity of the Kalman filter , the detection box , in the above formula, (u, v) are the central coordinate values of the traffic participant, that is, the abscissa and ordinate respectively, s is the area of the bounding box, r is the aspect ratio of the bounding box. Assuming that the aspect ratio r is constant, the other three variables are the derivatives of the corresponding respective variables with respect to time, w represents the width of the traffic participant, h represents the height h of the traffic participant, and c represents the detection confidence.

[0017] S22. Construct the state transition equation of the CV motion model; the expression of the state transition equation is:

[0018] ,

[0019] In the above formula, and represent the central coordinates of the traffic participant at time t + 1, and Represents the central coordinates of a traffic participant at time t. Represents the sensor acquisition interval.

[0020] S23. Construct a CV motion model according to the state transition equation, and obtain the predicted bounding boxes of all tracking IDs at the current time under the existing trajectories through Kalman filtering. The expression of Kalman filtering is as follows:

[0021]

[0022]

[0023] In the above formula, Represents that the Kalman filter calculates the predicted bounding boxes of all current tracking IDs at the current time using the state transition matrix and covariance matrix. Represents that the Kalman filter updates the parameters using the observed values, where Is the predicted state quantity at time t based on the state quantity at time t - 1. Is the state transition equation. Is the state quantity at time t - 1. Is the predicted covariance matrix at time t based on the covariance matrix at time t - 1. Is the covariance matrix at time t. Represents the process noise at time t. Is the Kalman gain at time t. Is the observation matrix at time t. Is the observation noise at time t. Is the state quantity at time t. Is the observed value at the moment. Is the covariance matrix at time t. Is the identity matrix.

[0024] Preferably, in step S3, it specifically includes the following steps:

[0025] S31. Obtain the detection bounding box at any time and the predicted bounding box at the current time.

[0026] S32. Calculate the IOU values of the detection bounding box at the current time and the predicted bounding boxes under different existing tracking IDs in turn. The calculation formula of the IOU value is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] In the above formula, A and B respectively represent the detection box and the prediction box, and respectively represent the abscissa and ordinate of the center point of object A, and respectively represent the width and height of object A, and respectively represent the abscissa and ordinate of the center point of object B, and respectively represent the width and height of object B, and respectively represent the horizontal and vertical overlap degrees of object A and object B, represents the overlapping area of object A and B, and respectively represent the areas of object A and B, represents the IOU value of object A and B, that is, the IOU value of the detection box and the prediction box;

[0034] S32. Construct a cost matrix based on the IOU values of all prediction boxes and detection boxes, set the values in the cost matrix that are less than the IOU threshold to infinity, and use the Hungarian algorithm for the cost matrix to obtain the optimal matching. Assign the tracking ID of the prediction box corresponding to the optimal matching to the detection box corresponding to the optimal matching;

[0035] S33. Assign a new tracking ID to the detection box whose IOU values are all less than the IOU threshold.

[0036] Preferably, in step S33, the tracking ID is represented by Arabic numerals, and the new tracking ID is the sum of the current maximum digital ID and a preset fixed value.

[0037] Preferably, in step S5, it specifically includes the following steps:

[0038] S51. Obtain the detection box at the current moment corresponding to the uncontinuously matched trajectory and the detection box with the same tracking ID that appeared last time, and mark this time period as the occlusion period;

[0039] S52. Construct a virtual trajectory of the traffic participant corresponding to the uncontinuously matched trajectory during the occlusion period; the virtual trajectory is constructed using the constant velocity interpolation algorithm, and the expression of the virtual trajectory is:

[0040]

[0041] In the above formula, Indicates the detection box when the tracking ID corresponding to the non - continuously matched trajectory of the traffic participant appeared last time. For the current moment and the detection box with the same tracking ID;

[0042] S53. Update the Kalman filter parameters during the occlusion period according to the virtual trajectory; the expression of the Kalman filter parameters to be updated is:

[0043]

[0044] In the above formula, Indicates the Kalman filter parameters to be updated during the occlusion period;

[0045] S54. Construct a virtual trajectory according to the detection boxes corresponding to the non - continuously matched trajectories before and after the occlusion period and by means of interpolation with a constant speed; the calculation formula for the position of the traffic participant on the virtual trajectory during the occlusion period is as follows:

[0046]

[0047]

[0048]

[0049]

[0050] In the above formula, and represent the abscissa and ordinate of the center point of the detection box at time t + 1, and represent the abscissa and ordinate of the center point of the detection box at time t, represents the sensor acquisition interval, and respectively represent the constant offset calculated by the division operation of the difference in the ordinate and abscissa of the center coordinates of the detection boxes before and after the occlusion period and the occlusion interval, and respectively represent and the derivatives with respect to time, representing the speed, and t represents the duration of the occlusion period.

[0051] This technical solution also provides a system for an autonomous driving tracking method in a long - time occlusion scenario. The system includes: characterized in that it includes a processor and a memory. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements any one of the autonomous driving tracking methods for a long - time occlusion scenario.

[0052] The technical solution also provides a computer storage medium for storing program data, which, when executed by a computer, implements an autonomous driving tracking method for long-time occlusion scenarios.

[0053] Compared with the prior art, the present invention provides an autonomous driving tracking method, system and storage medium for long-time occlusion scenarios, having the following beneficial effects:

[0054] 1. By constructing a virtual trajectory during the occlusion of a traffic participant and using the virtual trajectory to update the parameters of the Kalman filter again, the tracking ID does not jump when the traffic participant is redetected, improving the tracking performance of the autonomous driving system.

[0055] 2. The present invention can also improve the tracking performance of occluded targets in multiple technical fields such as autonomous driving, intelligent logistics, intelligent monitoring, data analysis, and behavior recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments and descriptions thereof are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings:

[0057] Figure 1 is a flowchart of the autonomous driving tracking method for long-time occlusion scenarios of the present invention;

[0058] Figure 2 is a schematic diagram showing the difference between the present invention and the prior tracking method;

[0059] Figure 3 is a schematic diagram of the virtual trajectory of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments. Thus, a full understanding of the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be obtained and implemented accordingly.

[0061] Those of ordinary skill in the art can understand that all or part of the steps of implementing the following embodiment methods can be completed by instructing relevant hardware through a program. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0062] To solve the problem that when a traffic participant is occluded for a long time, it is difficult for the existing DBT algorithm to accurately predict the trajectory of the traffic participant, the present invention provides an autonomous driving tracking method for long-term occlusion scenarios. The tracking method specifically includes the following steps:

[0063] S1. Obtain the detection results of all traffic participants at any moment based on the object detection algorithm, and assign a tracking ID to the detection results at the initial moment; the detection results include the detection box of the traffic participant, the center coordinates of the detection box, the size of the detection box, and the confidence level. For example, when the algorithm starts to run, a tracking ID needs to be assigned to the detection box to facilitate subsequent acquisition and further operations; the object detection algorithm includes mainstream object detection algorithms such as CenterNet, YOLO, and DINO, so as to better adapt to different scenarios. The present invention uses CenterNet as the object detection algorithm for comparison with other existing tracking methods. The performance comparison is shown in Table 1 below. It can be found that this indicates that the performance of the present invention in HOTA, MOTA, and AssA has been greatly improved. Compared with the existing tracking methods, as a tracker in the autonomous driving system, it can significantly improve the tracking performance.

[0064] S2. According to the trajectory data under different existing tracking IDs, use the Kalman filter to predict the prediction boxes of different existing tracking IDs at the current moment. The trajectory data is the tracking boxes of different tracking IDs at any moment. To further describe the process of calculating the prediction box through the Kalman filter, in step S2, it specifically includes the following steps:

[0065] S21. Define the state variables of the Kalman filter , the detection box , in the above formula, (u, v) are the center coordinate values of the traffic participant, s is the area of the bounding box, r is the aspect ratio of the bounding box. Assuming that the aspect ratio r is constant, the other three variables are the derivatives of the corresponding independent variables with respect to time, w represents the width of the traffic participant, h represents the height h of the traffic participant, and c represents the detection confidence level;

[0066] S22. Construct the state transition equation of the CV motion model; the expression of the state transition equation is:

[0067] ,

[0068] In the above formula, and represent the center coordinates of the traffic participant at time t + 1, and Denote the center coordinates of traffic participants at time t. Denote the sensor acquisition interval.

[0069] S23. Construct a CV motion model according to the state transition equation, and obtain the predicted bounding boxes of all tracking IDs at the current time under the existing trajectories through Kalman filtering. The expression of Kalman filtering is:

[0070]

[0071]

[0072] In the above formula, it means that the Kalman filter calculates the predicted bounding boxes of all current tracking IDs at the current time using the state transition matrix and covariance matrix, and represents that the Kalman filter updates the parameters using the observed values, where is the predicted state quantity at time t based on the state quantity at time t - 1. is the state transition equation. is the state quantity at time t - 1. is the predicted covariance matrix at time t based on the covariance matrix at time t - 1. is the covariance matrix at time t. represents the process noise at time t. is the Kalman gain at time t. is the observation matrix at time t. is the observation noise at time t. is the state quantity at time t. is the observed value at time, that is, the The calculation formulas for each item of the parameter. Thus, the calculation results can be substituted into the formulas for each item of the parameter to obtain the parameters of the predicted bounding box, that is , is the covariance matrix at time t. is the identity matrix.

[0073] S3. Calculate the IOU values of the detection bounding boxes at the current time and the predicted bounding boxes under different existing tracking IDs in turn, construct a cost matrix with the IOU values of all predicted bounding boxes and detection bounding boxes, set the values less than the IOU threshold in the cost matrix to infinity, and use the Hungarian algorithm on the cost matrix to obtain the optimal matching. Assign the tracking ID of the predicted bounding box corresponding to the optimal matching to the detection bounding box corresponding to the optimal matching, and assign a new tracking ID to the detection bounding box whose IOU values are all less than the IOU threshold. To further illustrate the process of assigning tracking IDs, in step S3, it specifically includes the following steps:

[0074] S31. Obtain the detection bounding boxes at any time and the predicted bounding boxes of all tracking IDs at the current time.

[0075] S32. Calculate the IOU values of the detection box at the current moment and the prediction boxes with different existing tracking IDs in sequence; the calculation formula of the IOU value is as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] In the above formula, A and B represent the detection box and the prediction box respectively, and represent the abscissa and ordinate of the center point of object A respectively, and represent the width and height of object A respectively, and represent the abscissa and ordinate of the center point of object B respectively, and represent the width and height of object B respectively, and represent the horizontal and vertical overlap degrees of object A and object B respectively, represents the overlapping area of object A and B, and represent the areas of object A and B respectively, represents the IOU value of object A and B, that is, the IOU value of the detection box and the prediction box;

[0083] S32. Set the values less than the IOU threshold in the cost matrix to infinity, and assign the tracking ID of the prediction box corresponding to the optimal match obtained by using the Hungarian algorithm for the cost matrix to the detection box corresponding to the optimal match; in actual use, the IOU threshold is generally set to 0.3.

[0084] S33. Assign a new tracking ID to the detection box whose IOU values are all less than the IOU threshold. In step S33, the tracking ID is represented by Arabic numerals, and the new tracking ID is the sum of the current maximum digital ID and a preset fixed value. The preset fixed value can be set to 1. For example, if the current maximum digital ID is 4, the new tracking ID will be assigned the value of 5.

[0085] S4. Determine whether all the tracking IDs except the new tracking ID at the current moment appeared at the previous moment;

[0086] If so, return to step S1;

[0087] If not, mark the trajectory where the detection box corresponding to the tracking ID that did not appear at the previous moment is located as an uncontinuously matched trajectory, that is, the traffic participant has been occluded. This tracking ID does not include a new tracking ID;

[0088] S5. According to the detection boxes of the traffic participant ID before and after occlusion, use the constant velocity interpolation method to construct the detection box at each moment during the occlusion period, so as to generate the virtual trajectory of the traffic participant during the occlusion period, and update the parameter values of the Kalman filter to which the traffic participant belongs during the occlusion period to ensure that the traffic participant can be continuously tracked subsequently. In step S5, it specifically includes the following steps:

[0089] S51. Obtain the detection box at the current moment corresponding to the uncontinuously matched trajectory and the detection box with the same tracking ID that appeared last time, and mark this time period as the occlusion period;

[0090] S52. Construct the virtual trajectory of the traffic participant corresponding to the uncontinuously matched trajectory during the occlusion period; the virtual trajectory is constructed using the constant velocity interpolation algorithm, and the expression of the virtual trajectory is:

[0091]

[0092] In the above formula, represents the detection box when the tracking ID of the traffic participant corresponding to the uncontinuously matched trajectory appeared last time, is the detection box with the same tracking ID as the detection box at the current moment;

[0093] S53. Update the Kalman filter parameters during the occlusion period according to the virtual trajectory; the expression of the Kalman filter parameters that need to be updated is:

[0094]

[0095] In the above formula, represents the Kalman filter parameters that need to be updated during the occlusion period; is the observed value of the virtual trajectory of the traffic participant during the occlusion period, so as to calculate the predicted value at any moment on the virtual trajectory.

[0096] S54. According to the detection boxes corresponding to the uncontinuously matched trajectory before and after the occlusion period, and use the interpolation method of constant velocity to construct the virtual trajectory; the calculation formula for the position of the traffic participant on the virtual trajectory during the occlusion period is as follows:

[0097]

[0098]

[0099]

[0100]

[0101] In the above formula, and represent the abscissa and ordinate of the center point of the detection box at time t + 1, and represent the abscissa and ordinate of the center point of the detection box at time t, represents the sensor acquisition interval, and respectively represent the constant offset calculated by the division operation of the position differences of the ordinate and abscissa of the center coordinates of the detection box before and after the occlusion period and the occlusion interval, and respectively represent and The derivatives with respect to time represent velocity, and t represents the duration of the occlusion period.

[0102] As Figure 2 shown, the red box is the detection box, the orange box is the movement trajectory of the traffic participant, the blue box is the trajectory where tracking is lost, and the blue dashed box is the tracking trajectory predicted by the Kalman filter. At time t + 1, the first traffic participant is occluded, and at time t + 2, the first traffic participant is tracked. In the existing method, the matching degree between the tracking box and the actual traffic participant is very small, resulting in subsequent tracking loss. The method in the present invention ensures that subsequent tracking of the first traffic participant will not be lost by recalculating the KF parameters from time t to time t + 2.

[0103] As Figure 3 shown, in figure (a), at time t1, the traffic participant is lost due to occlusion, and at time t2, the tracking box of the traffic participant trajectory predicted by KF is re-associated with the detection box with the same tracking ID as before. In figure (b) for the existing method, at the next moment after re-association, even though the KF state is updated at time t2, since there is still a direction difference between the true traffic participant trajectory and the KF predicted trajectory, tracking will be lost again (blue) at the next moment after t2. In figure (c) for the virtual trajectory (red) of the traffic participant occlusion period proposed by the present invention, it can be seen from the figure that the virtual trajectory is closer to the true state observation of the object, and the detection box with the previously appeared tracking ID can still be associated at the moments after t2.

[0104] The technical solution also provides a system for an autonomous driving tracking method for a long-time occlusion scenario, the system including: a processor and a memory, the memory being used for storing a computer program, and the computer program, when executed by the processor, implementing an autonomous driving tracking method for a long-time occlusion scenario.

[0105] The technical solution also provides a computer storage medium, the computer storage medium being used for storing program data, and the program data, when executed by a computer, being used to implement an autonomous driving tracking method for a long-time occlusion scenario.

[0106] The above embodiments have introduced the present invention in detail. Specific examples are used in this article to elaborate on the principle and embodiments of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An automatic driving tracking method for a long-time occlusion scenario, characterized in that, The tracking method specifically includes the following steps: S1. Obtain the detection results of all traffic participants at any moment based on the object detection algorithm, and assign a tracking ID to the detection results at the initial moment; the detection results include the detection box of the traffic participant, the center coordinates of the detection box, the size of the detection box, and the confidence level; S2. According to the trajectory data under different existing tracking IDs, use the Kalman filter to predict the prediction boxes of different existing tracking IDs at the current moment; S3. Calculate the IOU values of the detection box at the current moment and the prediction boxes under different existing tracking IDs in turn, and form an IOU cost matrix. Set the values less than the IOU threshold in the cost matrix to infinity. Use the Hungarian algorithm on the cost matrix to obtain the optimal matching. Assign the tracking ID of the prediction box corresponding to the optimal matching to the detection box corresponding to the optimal matching, and assign a new tracking ID to the detection box whose IOU values are all less than the IOU threshold; S4. Determine whether all the tracking IDs except the new tracking ID at the current moment appeared in the previous moment; If so, return to step S1; If not, mark the trajectory where the detection box corresponding to the tracking ID that did not appear in the previous moment is located as an uncontinuously matched trajectory, and this tracking ID does not include the new tracking ID; S5. According to the detection boxes of the tracking ID corresponding to the uncontinuously matched trajectory at the current moment and the last appearance, use the fixed-speed interpolation algorithm to generate virtual trajectories during the occlusion period to update the parameter values of the Kalman filter during the occlusion period.

2. The tracking method according to claim 1, characterized in that In step S1, the object detection algorithm is the CenterNet algorithm.

3. The tracking method according to claim 1, wherein In step S2, it specifically includes the following steps: S21. Define the state variables of the Kalman filter , the detection box , in the above formula, (u, v) are the central coordinate values of the traffic participant, that is, the abscissa and ordinate respectively, s is the area of the bounding box, r is the aspect ratio of the bounding box. Assuming that the aspect ratio r is constant, the other three variables are the derivatives of the corresponding independent variables with respect to time, w represents the width of the traffic participant, h represents the height h of the traffic participant, and c represents the detection confidence; S22. Construct the state transition equation of the CV motion model; the expression of the state transition equation is: , In the above formula, and represent the central coordinates of the traffic participant at time t + 1, and represent the central coordinates of the traffic participant at time t, represents the sensor acquisition interval; S23. Construct the CV motion model according to the state transition equation, and obtain the prediction boxes of all existing trajectories of tracking IDs at the current moment through Kalman filtering. The expression of Kalman filtering is: In the above formula, indicates that the Kalman filter calculates the predicted bounding boxes at the current time for all current tracking IDs using the state transition matrix and covariance matrix. indicates that the Kalman filter updates the parameters using the observed values, where is the predicted state quantity at time t based on the state quantity at time t - 1. is the state transition equation. is the state quantity at time t - 1. is the predicted covariance matrix at time t based on the covariance matrix at time t - 1. is the covariance matrix at time t. represents the process noise at time t. is the Kalman gain at time t. is the observation matrix at time t. is the observation noise at time t. is the state quantity at time t. is the observed value at the moment. is the covariance matrix at time t. is the identity matrix.

4. The tracking method according to claim 1, characterized in that In step S3, it specifically includes the following steps: S31. Obtain the detection box at any moment and the prediction box at the current moment; S32. Calculate the IOU values of the detection box at the current moment and the prediction boxes under different existing tracking IDs in turn; the calculation formula of the IOU value is as follows: In the above formula, A and B represent the detection box and the prediction box respectively. and represent the abscissa and ordinate of the center point of object A respectively. and represent the width and height of object A respectively. and represent the abscissa and ordinate of the center point of object B respectively. and represent the width and height of object B respectively. and represent the horizontal and vertical overlap degrees of object A and object B respectively. represents the overlapping area of object A and B. and represent the areas of object A and B respectively. represents the IOU value of object A and B, that is, the IOU value of the detection box and the prediction box. S32. Construct a cost matrix according to the IOU values of all prediction boxes and detection boxes. Set the values less than the IOU threshold in the cost matrix to infinity. Use the Hungarian algorithm on the cost matrix to obtain the optimal matching. Assign the tracking ID of the prediction box corresponding to the optimal matching to the detection box corresponding to the optimal matching; S33. Assign a new tracking ID to the detection box whose IOU values are all less than the IOU threshold.

5. The tracking method according to claim 4, wherein In step S33, the tracking ID is represented by Arabic numerals, and the new tracking ID is the sum of the largest existing digital ID and a preset fixed value.

6. The tracking method according to claim 1, wherein In step S5, it specifically includes the following steps: S51. Obtain the detection box at the current moment corresponding to the uncontinuously matched trajectory and the detection box of the same tracking ID at the last appearance, and mark this time period as the occlusion period; S52. Construct a virtual trajectory of the traffic participant corresponding to the un - continuously matched trajectory during the occlusion period; the virtual trajectory is constructed using a constant - velocity interpolation algorithm, and the expression of the virtual trajectory is: In the above formula, represents the detection box of the traffic participant corresponding to the non - continuously - matched trajectory at the time of its last appearance, is the detection box at the current moment with the same tracking ID as the detection box ; S53. Update the Kalman filter parameters during the occlusion period according to the virtual trajectory; the expression of the Kalman filter parameters to be updated is: In the above formula, represents the Kalman filter parameters that need to be updated during the occlusion period; S54. Construct a virtual trajectory by using the detection frames corresponding to the un - continuously matched trajectories before and after the occlusion period and in an interpolation manner with a constant velocity; the calculation formula for the position of the traffic participant on the virtual trajectory during the occlusion period is as follows: In the above formula, and represent the abscissa and ordinate of the center point of the detection box at time t+1, and represent the abscissa and ordinate of the center point of the detection box at time t, represents the sensor acquisition interval, and respectively represent the constant offsets calculated by the division operation of the position differences of the ordinate and abscissa of the center coordinates of the detection box before and after the occlusion period and the occlusion interval, and respectively represent and the derivatives with respect to time, representing the velocity, and t represents the duration of the occlusion period.

7. A system for implementing the tracking method according to any one of claims 1-6, characterized in that, It includes a processor and a memory. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the tracking method according to any one of claims 1 - 6.

8. A computer storage medium, which is used to store program data, and when the program data is executed by a computer, it implements the autonomous driving tracking method for a long - time occlusion scenario according to any one of claims 1 - 7.

Images