A multi-target tracking method in a traffic scene

Abstract

The application provides a multi-target tracking method in a traffic scene, and point tracks output by a radar are clustered to form clusters; information of the clusters comprises average positions of radar point tracks in the clusters, average speeds, numbers of radar point tracks in the clusters and average signal-to-noise ratios; when track updating is performed, position and speed information of all associated clusters, a proportion of a product of the average signal-to-noise ratio of the radar point tracks in the current cluster and the number of the radar point tracks in the current cluster in a total product of the average signal-to-noise ratios of the radar point tracks in all associated clusters and the number of the radar point tracks in all associated clusters are weighted, and the weighted result is taken as an input of Kalman filtering; the application can effectively alleviate the problem of track splitting of large-size targets in target tracking, and is suitable for multi-target tracking in a complex traffic scene.

Description

Technical Field

[0001] This invention relates to the field of radar data processing, and in particular to a multi-target tracking method in traffic scenarios. Background Technology

[0002] Compared to sensors such as optical cameras and lidar, millimeter-wave radar offers the advantage of all-day, all-weather operation, playing an irreplaceable role in the field of intelligent transportation. Target tracking based on millimeter-wave radar data provides crucial traffic target information for intelligent transportation systems.

[0003] Target tracking in traffic scenarios mainly involves two techniques: target tracking preprocessing (clustering) and target tracking. In traffic scenarios, the density-based DBSCAN clustering algorithm is commonly used. DBSCAN has advantages such as not requiring a pre-defined number of clusters and high real-time performance. In DBSCAN, a threshold is needed to determine whether radar points are core samples. Similar clustering algorithms require a threshold to determine if they represent the same target. In complex traffic scenarios, setting this threshold is difficult to balance large and small targets simultaneously, thus complicating subsequent target tracking. If the threshold is set too high, small, closely spaced targets may be clustered together; if it is set too low, large targets may be clustered into multiple clusters.

[0004] Traditional target tracking methods use the cluster closest to the current track as input for the Kalman filter when filtering the current track. However, if the initial clustering separates large targets into multiple clusters, this can lead to track splitting issues.

[0005] Existing methods for solving the track splitting problem include: methods based on optimized track management, but these do not address the problem at its core, failing to improve the accuracy of the filtering and tracking results, and thus have limited effectiveness in optimizing tracks; replacing traditional intra-frame clustering with inter-frame clustering, but this involves setting many parameters, such as the number of spatial segments and the number of merged frames. These parameter settings vary depending on the scenario and the target signal-to-noise ratio, and are often based on experience; and combining iterative processes to optimize measurements, but multiple iterations result in poor real-time performance in complex multi-target traffic application scenarios.

[0006] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-target tracking method in traffic scenarios, which can effectively solve the problem of track splitting of large-sized targets in complex traffic scenarios and improve the accuracy of target tracking results.

[0008] To achieve the above objectives, the present invention provides a multi-target tracking method in traffic scenarios, comprising the following steps:

[0009] Step S1: Cluster the radar spot data to form clusters;

[0010] Step S2: Determine if the track has started successfully;

[0011] Step S3: For each track, search for clusters within the relevant gates to obtain the associated clusters;

[0012] Step S4: When updating the track, the position and velocity information of all clusters associated with the current track are weighted by the ratio of the product of the average signal-to-noise ratio of radar points in the current cluster and the number of radar points to the sum of the products of the average signal-to-noise ratio of radar points in all associated clusters. The weighted result is used as the input of the Kalman filter.

[0013] Step S5: Perform Kalman filtering on the weighted intra-cluster lateral position, longitudinal position, and velocity.

[0014] The cluster information includes: the average lateral position, average longitudinal position, average velocity, number of radar points within the cluster, and average signal-to-noise ratio of radar points within the cluster.

[0015] An intuitive method is used to determine whether a trajectory has started successfully. Trajectory information includes: position, velocity, whether there is a target associated with it from the current frame to the previous M frames, and the number of consecutive frames where association failed in the current frame and previous frames. Threshold conditions include: whether the target's velocity is within the specified range [v]. min ,v max Within ], is the target acceleration within the range of values ​​[a]? min ,a max [Inside, where v] min v max a min a max These represent the minimum speed, maximum speed, minimum acceleration, and maximum acceleration that the target may achieve, respectively. If q2 frames of data in consecutive q1 frames meet the threshold condition, the trajectory is considered to have started successfully.

[0016] Similarity is used to set the correlation gate, and the absolute values ​​of the position difference Δr and velocity difference Δv between the cluster and the track are calculated. If Δr is less than the set threshold ρ, the cluster is considered closed. r And Δv is less than the set threshold ρ v If so, then the cluster is considered to be associated with the current track.

[0017] Kalman filtering involves state prediction and updating. The state prediction formula is as follows:

[0018]

[0019]

[0020] Where the subscripts k-1 and k represent the previous time and the current time, respectively. This represents the posterior state estimate from the previous time step. Let P represent the prior state estimate of the current state obtained using prior knowledge, A represent the transition matrix from the previous state to the current state, and P represent the prior state estimate. k - Let P represent the error covariance matrix of the prior state estimation. k-1 Let represent the error covariance matrix of the posterior state estimate at the previous time step, and Q represent the covariance matrix of the process noise.

[0021] The state update formula is:

[0022]

[0023]

[0024]

[0025] Among them, K k Let H represent the Kalman filter gain, H represent the transformation matrix from the current state to the measurement, R represent the covariance matrix of the measurement noise, and z represent the Kalman filter gain. k This represents the measurement value at the current time. Let I represent the posterior state estimate of the current state, and let P represent the identity matrix. k Let represent the error covariance matrix of the posterior state estimate.

[0026] The multi-target tracking method further includes: step S6, for each track, determining whether there is no measurement data associated with the track for N consecutive frames; if so, the track is terminated; otherwise, the track update continues.

[0027] This invention, when clustering radar traces, outputs not only the average position and average velocity of radar traces within a cluster, but also the number of radar traces within the cluster and the average signal-to-noise ratio (SNR). During track updates, the position and velocity of all clusters within the gate are weighted proportionally to the product of the average SNR and number of radar traces within the current cluster, relative to the sum of the products of the average SNR and number of radar traces within all associated clusters. This weighted average SNR of these products is then used as input for Kalman filtering. This ensures that even when a large target is clustered into multiple clusters, the corresponding clusters can still be correctly associated with their respective tracks. If a cluster is noisy, its weight is small due to the low number of radar traces and low SNR, thus having little impact on the weighted result. This invention effectively solves the problem of track fragmentation for large targets in complex traffic scenarios, improving the accuracy of target tracking results. Attached Figure Description

[0028] Figure 1 This is a flowchart of a multi-target tracking method in a traffic scenario provided by the present invention.

[0029] Figure 2 This is a real-world traffic scenario for data collection.

[0030] Figure 3 It is the lateral distance of the target tracking trajectory before optimization.

[0031] Figure 4 It is the longitudinal distance of the target tracking trajectory before optimization.

[0032] Figure 5 It is the lateral distance of the optimized target tracking trajectory.

[0033] Figure 6 It is the longitudinal distance of the optimized target tracking trajectory. Detailed Implementation

[0034] The following is based on Figures 1-6 The preferred embodiments of the present invention will be described in detail below.

[0035] To address the issue that radar spot clustering threshold settings cannot simultaneously accommodate both large and small targets, potentially resulting in the same large target being clustered into multiple clusters and thus affecting target tracking, this invention provides a multi-target tracking method for traffic scenarios, such as... Figure 1 As shown, it includes the following steps:

[0036] Step S1: Cluster the radar point data to form clusters. The information of the cluster includes: the average lateral position, average longitudinal position, average velocity, number of radar points in the cluster, and average signal-to-noise ratio of radar points in the cluster.

[0037] Step S2: Determine whether the track has started successfully.

[0038] A visual method can be used to determine whether a flight path has started successfully. If q2 frames out of q1 consecutive frames meet the threshold condition, the flight path is considered to have started successfully. Based on actual scenarios and experience, in this embodiment, q1 = 5 and q2 = 3.

[0039] The trajectory information includes: position, velocity, whether there is a target associated with it from the current frame to the previous M frames, and the number of consecutive frames where association failed in the current frame and previous frames. The intuitive method specifically involves determining whether the target's velocity is within the range [v]. min ,v max [a] Whether the target acceleration is within the range of values. min ,a max [Inside.] Among them, v minv max a min a max These represent the minimum speed, maximum speed, minimum acceleration, and maximum acceleration that the target can achieve, respectively.

[0040] Step S3: For each track, search for clusters within the relevant gates of the track to obtain the associated clusters.

[0041] The relevant gate can be set using similarity, that is, by calculating the absolute value of the position difference Δr and the absolute value of the velocity difference Δv between the cluster and the track. If Δr is less than the set threshold ρ, the gate is selected. r And Δv is less than the set threshold ρ v If so, then the cluster is considered to be associated with the current track.

[0042] Step S4: When updating the track, the position and velocity information of all clusters associated with the current track are weighted by the ratio of the product of the average signal-to-noise ratio of radar points in the current cluster and the number of radar points to the sum of the products of the average signal-to-noise ratio of radar points in all associated clusters. The weighted result is used as the input of the Kalman filter.

[0043] Step S5: Perform Kalman filtering on the weighted intra-cluster lateral position, longitudinal position, and velocity.

[0044] Kalman filtering involves state prediction and updating. The state prediction formula is as follows:

[0045]

[0046] P k - =AP k-1 A T +Q

[0047] Where the subscripts k-1 and k represent the previous time and the current time, respectively. This represents the posterior state estimate from the previous time step. Let P represent the prior state estimate of the current state obtained using prior knowledge, A represent the transition matrix from the previous state to the current state, and P represent the prior state estimate. k - Let P represent the error covariance matrix of the prior state estimation. k-1 Let represent the error covariance matrix of the posterior state estimate at the previous time step, and Q represent the covariance matrix of the process noise.

[0048] The state update formula is:

[0049] K k =P k - H T HP k- H T +R) -1

[0050]

[0051]

[0052] Among them, K k Let H represent the Kalman filter gain, H represent the transformation matrix from the current state to the measurement, R represent the covariance matrix of the measurement noise, and z represent the Kalman filter gain. k This represents the measurement value at the current time. Let I represent the posterior state estimate of the current state, and let P represent the identity matrix. k Let represent the error covariance matrix of the posterior state estimate.

[0053] Step S6: For each track, determine whether there is no measurement data associated with the track for N consecutive frames. If so, end the track. If not, continue track updating.

[0054] Example:

[0055] Step 1: Cluster the radar spot data for each frame to form clusters. Cluster information includes the cluster's lateral position x, vertical position y, velocity v, signal-to-noise ratio (SNR) snr, and the number of radar spots within the cluster n. The lateral position x of a cluster is the average lateral position of the radar spots within the cluster. Assuming P radar spots are clustered into one cluster, the lateral distance of that cluster is... Where, x i ′,i=1…P represents the lateral distance of the i-th radar trace. Similarly, the longitudinal position y of the cluster is the average longitudinal position of the radar traces within the cluster, the velocity v of the cluster is the average velocity of the radar traces within the cluster, and the signal-to-noise ratio snr of the cluster is the average signal-to-noise ratio of the radar traces within the cluster on the range-velocity map.

[0056] Step 2: Determine the start of the track. The start of the track can be determined intuitively: if q2 frames out of q1 consecutive frames meet a threshold condition, the track is considered to have started successfully. The track information includes: position, speed, whether there is a target associated with the target from the current frame to the previous 5 frames, and the number of consecutive frames where association failed (including the current frame and previous frames).

[0057] Step 3: Search for clusters within the relevant gate of the track to obtain the associated clusters. The relevant gate can be set using similarity, that is, calculate the position difference and velocity difference between the cluster and the track. If both the position difference and velocity difference are less than the set threshold, the cluster is considered to be associated with the current track.

[0058] Step 4: When updating the track, the position and velocity of all clusters within the gate are weighted by the ratio of the product of the average signal-to-noise ratio of radar traces in the current cluster and the number of radar traces to the sum of the products of the average signal-to-noise ratio of radar traces in all associated clusters, and the input of the Kalman filter is obtained.

[0059] Suppose the current track has M associated clusters, and the lateral position of the M associated clusters is x. i (i = 1…M), with vertical position y i (i = 1…M), velocity v i (i = 1…M), signal-to-noise ratio is snr i (i = 1…M), the number of radar points within the cluster is n. i (i = 1…M);

[0060] The input to the Kalman filter is:

[0061]

[0062]

[0063]

[0064] Wherein, weight α i For the first i The average signal-to-noise ratio (SNR) of radar traces within a cluster. i With the number of radar points n i The product of these factors accounts for the proportion of the sum of the products of the average signal-to-noise ratio of radar traces within all associated clusters and the number of radar traces, i.e.:

[0065]

[0066] Step 5: Perform Kalman filtering based on the track information and Kalman filter input.

[0067] Step 6: Determine if the track termination condition is met. If yes, terminate the track. If not, continue updating the track.

[0068] This embodiment demonstrates target tracking processing in a real-world traffic scenario. Target tracking was performed using both conventional and the method of this invention, employing the same radar data. The millimeter-wave radar frame period was 0.1 seconds, and a total of 130 frames were collected. The conventional target tracking method proceeds as follows: Radar point data is clustered using the DBSCAN algorithm to obtain clusters; track initiation is determined; during track updates, the cluster closest to the relevant gate is used as the associated object and is input to a linear Kalman filter; the condition for track termination is determined, and if yes, the track ends; otherwise, track updates continue. The clustering operation of this invention also uses the DBSCAN algorithm. Cluster information includes not only the position and velocity of conventional clusters but also the number of radar points within the cluster and the average signal-to-noise ratio (SNR). During track updates, the position and velocity of all clusters within the gate are weighted proportionally to the product of the average SNR and number of radar points within the current cluster, relative to the sum of the products of the average SNR and number of radar points within all associated clusters, and this weighted average is used as input to the Kalman filter.

[0069] like Figure 2 The image shows the actual traffic scene at frame 70 of the data acquisition. A large truck is located approximately 200 meters in the middle lane. The radar plane is perpendicular to the road's direction, illuminating the area ahead. Let the target's coordinate system be: the origin is the radar antenna phase center; the positive Y-axis (longitudinal) is the direction perpendicular to the radar plane and extending towards the road; the X-axis (lateral) is along the radar plane; and the positive Z-axis is the direction perpendicular to both the X and Y axes, pointing upwards. The positive X-axis is determined using the right-hand coordinate system rule. Figure 3 This is the lateral trajectory of the target tracking without the optimization of this invention. In the figure, the horizontal axis represents the frame, and the vertical axis represents the lateral trajectory (X-axis trajectory). Figure 4 The figures show the longitudinal trajectory of the target tracking without optimization by this invention. The horizontal axis represents the frame, and the vertical axis represents the longitudinal trajectory (Y-axis trajectory). Both figures use points to mark the clustering results and circles to mark the target tracking results. It can be seen from the figures that the radar point output has a large error compared to the X-position information, but... Figure 4 It can be seen that the longest longitudinal trajectory exhibits track splitting issues (represented by hollow rings) in frames 38–78 and 85–100. Figure 5 and Figure 6 The target tracking results obtained using the method of this invention do not have the problem of track splitting. It can be seen that this invention can effectively alleviate the problem of track splitting of large targets in traffic scenarios and is suitable for multi-target tracking in complex traffic scenarios.

[0070] It should be noted that, in the embodiments of the present invention, the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the embodiments. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0071] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0072] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A multi-target tracking method in a traffic scenario, characterized in that, Includes the following steps: Step S1: Cluster the radar spot data to form clusters; Step S2: Determine if the track has started successfully; Step S3: For each track, search for clusters within the relevant gates to obtain the associated clusters; Step S4: When updating the track, the position and velocity information of all clusters associated with the current track are weighted by the ratio of the product of the average signal-to-noise ratio of radar points in the current cluster and the number of radar points to the sum of the products of the average signal-to-noise ratio of radar points in all associated clusters. The weighted result is used as the input of the Kalman filter. Step S5: Perform Kalman filtering on the weighted intra-cluster lateral position, longitudinal position, and velocity.

2. The multi-target tracking method in a traffic scenario as described in claim 1, characterized in that, The cluster information includes: the average lateral position, average longitudinal position, average velocity, number of radar points within the cluster, and average signal-to-noise ratio of radar points within the cluster.

3. The multi-target tracking method in a traffic scenario as described in claim 1, characterized in that, An intuitive method is used to determine whether a trajectory has started successfully. Trajectory information includes: position, velocity, whether there is a target associated with it from the current frame to the previous M frames, and the number of consecutive frames where association failed in the current frame and previous frames. Threshold conditions include: whether the target's velocity is within the specified range [v]. min v max Within, is the target acceleration...? Within the range of values ​​[a] min ,a max [Inside, where v] min v max a min a max These represent the minimum speed, maximum speed, minimum acceleration, and maximum acceleration that the target may achieve, respectively. If q2 frames of data in consecutive q1 frames meet the threshold condition, the trajectory is considered to have started successfully.

4. The multi-target tracking method in a traffic scenario as described in claim 1, characterized in that, Similarity is used to set the correlation gate, and the absolute values ​​of the position difference Δr and velocity difference Δv between the cluster and the track are calculated. If Δr is less than the set threshold ρ, the cluster is considered closed. r And Δv is less than the set threshold ρ v If so, then the cluster is considered to be associated with the current track.

5. The multi-target tracking method in a traffic scenario as described in claim 1, characterized in that, Kalman filtering involves state prediction and updating. The state prediction formula is as follows: Where the subscripts k-1 and k represent the previous time and the current time, respectively. This represents the posterior state estimate from the previous time step. Let P represent the prior state estimate of the current state obtained using prior knowledge, A represent the transition matrix from the previous state to the current state, and P represent the prior state estimate. k - Let P represent the error covariance matrix of the prior state estimation. k-1 Let represent the error covariance matrix of the posterior state estimate at the previous time step, and Q represent the covariance matrix of the process noise. The state update formula is: Among them, K k Let H represent the Kalman filter gain, H represent the transformation matrix from the current state to the measurement, R represent the covariance matrix of the measurement noise, and z represent the Kalman filter gain. k This represents the measurement value at the current time. Let I represent the posterior state estimate of the current state, and let P represent the identity matrix. k Let represent the error covariance matrix of the posterior state estimate.

6. The multi-target tracking method in a traffic scenario as described in claim 1, characterized in that, The multi-target tracking method further includes: step S6, for each track, determining whether there is no measurement data associated with the track for N consecutive frames; if so, the track is terminated; otherwise, the track update continues.

Images