A multi-vehicle tracking method based on a millimeter wave radar under a traffic scene

Abstract

The application relates to a kind of millimeter wave radar multi-vehicle tracking methods based on traffic scene, comprising: the original echo data of millimeter wave radar is preprocessed, and a plurality of frames of point cloud data are obtained;Each frame of point cloud data is clustered using a hierarchical clustering algorithm based on target information;Rectangular association wave door is set in combination with standard lane width and safe following distance, and the rectangular association wave door is corrected according to the historical information obtained by target tracking;All target centroid points in the corrected rectangular association wave door are fused according to probability;The existing track associated with the equivalent centroid point and the existing track not associated with the equivalent centroid point are updated, and a plurality of updated tracks are obtained;For the target centroid point not assigned to the existing track, the logical method is used for track initiation, and a plurality of stable starting tracks are obtained;The target tracking track is obtained by managing the track. The method combines the characteristics of radar measurement distribution and motion mode of vehicle target, and realizes the robust tracking of multi-vehicle target.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, specifically relating to a millimeter-wave radar multi-vehicle tracking method based on traffic scenarios. Background Technology

[0002] In recent years, the field of intelligent transportation has developed rapidly, and the application of intelligent transportation equipment has made travel planning more convenient. Currently, the main traffic monitoring methods include millimeter-wave radar and cameras. Millimeter-wave radar can detect and track vehicle targets, while cameras can accurately capture relevant information about vehicle targets. Compared with cameras, millimeter-wave radar has significant advantages, such as a wider monitoring range, longer distance, ease of achieving beyond-line-of-sight perception, smaller data upload volume reducing system load, and, more importantly, strong anti-interference capabilities, unaffected by weather and environment. Therefore, millimeter-wave radar data processing technology is an important aspect driving the development of intelligent transportation, and multi-target tracking methods based on millimeter-wave radar are a key technology that presents significant challenges and difficulties in millimeter-wave radar perception.

[0003] In traffic scenarios, millimeter-wave radar multi-target tracking technology plays a crucial role in millimeter-wave radar perception. Compared to single-measurement millimeter-wave radar, multi-target tracking ensures higher accuracy and reliability by providing continuous position and velocity information for multiple targets. Furthermore, it can effectively eliminate false detections. Due to the increased resolution of millimeter-wave radar, vehicle targets typically span multiple resolution cells; therefore, the core challenge of multi-target tracking lies in clustering and data association.

[0004] In recent years, research on clustering and data association methods for multi-target tracking in traffic scenarios has become increasingly popular. Existing clustering and data association methods have achieved good tracking performance in simple millimeter-wave radar multi-target tracking tasks, but they still have shortcomings: existing clustering and data association methods do not consider improving the clustering and association stages by combining the target measurement distribution characteristics and motion patterns. When faced with multi-target tracking tasks in complex environments, their performance will degrade. For example, in traffic scenarios, due to the large number of vehicles and their proximity, it is difficult to achieve good tracking results by directly applying the above methods to multi-vehicle target tracking tasks. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a millimeter-wave radar multi-vehicle tracking method based on traffic scenarios. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a millimeter-wave radar multi-vehicle tracking method based on traffic scenarios, characterized by the following steps:

[0007] The raw echo data from the millimeter-wave radar is preprocessed to obtain several frames of point cloud data.

[0008] A hierarchical clustering algorithm based on target information is used to cluster each frame of point cloud data, and the target centroid of each cluster is calculated.

[0009] Using the predicted point of the existing track as the center, a rectangular associated gate is set in combination with the standard lane width and safe following distance. The rectangular associated gate is corrected based on the historical information obtained from target tracking, and the center position of the corrected rectangular associated gate is predicted.

[0010] Based on the aforementioned center position, a probabilistic data association algorithm is used to fuse all target centroids within the corrected rectangular associated gate according to probability, thereby obtaining an equivalent centroid.

[0011] For existing tracks associated with the equivalent centroid, the standard Kalman filter algorithm is used to update the tracks, and for existing tracks not associated with the equivalent centroid, the predicted values ​​in the standard Kalman filter algorithm are used to update the tracks, resulting in several updated tracks.

[0012] For target centroids that have not been assigned to existing tracks, a logical method is used to initiate the track, resulting in several stable initial tracks.

[0013] The updated track or the stable starting track is terminated according to the termination condition, and noise is judged on the terminated track to obtain the target tracking track.

[0014] In one embodiment of the present invention, the raw echo data of the millimeter-wave radar is preprocessed to obtain several frames of point cloud data, including:

[0015] The raw echo data from the millimeter-wave radar is sequentially mixed, low-pass filtered, and sampled at discrete time to obtain a discrete-time signal:

[0016]

[0017]

[0018] in, For the first Difference frequency signal in discrete-time form of a single channel The amplitude of the difference frequency signal. The phase of the difference frequency signal. For discrete-time angular frequency, The number of discrete-time sampling points. Number of transmitting antennas Number of receiving antennas The frequency of the difference frequency signal. The discrete-time sampling frequency;

[0019] The discrete-time signals are arranged sequentially according to the number of sampling points in the fast time dimension, the number of sampling points in the slow time dimension, and the number of antenna channels to obtain a three-dimensional signal:

[0020]

[0021] in, For the number of sampling points in the fast time dimension, The number of sampling points in the slow time dimension. This refers to the number of antenna channels;

[0022] Perform discrete-time Fourier transforms on the three-dimensional signal sequentially according to the fast time dimension, slow time dimension, and antenna dimension to obtain the range-Doppler-angle three-dimensional tensor:

[0023]

[0024] in, This represents the distance-Doppler-angle tensor. This represents the discrete-time Fourier transform. This represents the number of points in the fast-time discrete-time Fourier transform. This represents the number of points in the slow-time discrete-time Fourier transform. The number of discrete-time Fourier transform points representing the antenna channel dimension;

[0025] Peak detection is performed on the distance-Doppler-angle three-dimensional tensor to obtain target state information:

[0026]

[0027] Here, CFAR stands for Constant False Alarm Rate (CFAR) detection method. Indicates the distance to the target. Indicates the Doppler velocity of the target. Indicates the target's azimuth angle;

[0028] The target state information is transformed to obtain point cloud data for each frame:

[0029]

[0030] in, The target point has two-dimensional coordinates. The velocity at the target point.

[0031] In one embodiment of the present invention, a hierarchical clustering algorithm based on target information is used to cluster each frame of point cloud data, and the target centroid of each cluster is calculated, including:

[0032] Input clustering parameters ,in, Parameters for target information, To minimize the number of points in the cluster, the target information is expressed as a rectangular neighborhood;

[0033] Select any core sample point without a category from each frame of point cloud data, and find all sample sets that the core sample point can reach in density according to the clustering parameters to form a cluster. Repeat the clustering until all core sample points have a category to obtain several first point trace clusters.

[0034] When the velocity variance of all points in the first point cluster is greater than the first threshold, the new clustering parameters are used to cluster all points in the first point cluster to obtain several second point clusters.

[0035] When the distance between the center points of two second point trace clusters is less than the second threshold and the velocity variance is less than the third threshold, the two second point trace clusters are merged to obtain several clusters.

[0036] When there are multiple core sample points in the cluster, the multiple core sample points are weighted according to the magnitude of each point to obtain the target centroid point; when there is only one core sample point in the cluster, the core sample point is used as the target centroid point.

[0037] In one embodiment of the present invention, correcting the rectangular associated gate based on historical information obtained from target tracking and predicting the center position of the corrected rectangular associated gate includes:

[0038] For the existing track, calculate the current time before... The average velocity vector of the velocity vector in the frame track is used as the target heading angle, and the rectangular associated gate is rotated according to the target heading angle to obtain the corrected rectangular associated gate;

[0039] The center position is obtained by predicting the corrected rectangular associated gate using the average velocity vector and the position information of the target state at the previous moment.

[0040] In one embodiment of the present invention, based on the center position, a probabilistic data association algorithm is used to fuse all target centroids within the corrected rectangular associated gate according to probability to obtain equivalent centroids, including:

[0041] Calculate the association probability of all target centroids within the corrected rectangular associated gate based on the center position:

[0042]

[0043] in, This represents the probability of detecting the target. This represents the probability that the target appears within the corrected rectangular correlation gate. express The probability originating from the target, This indicates the centroid point that falls into the corrected rectangular correlated gate. Indicates the predicted center location. Describing covariance, Indicates clutter density;

[0044] Based on the correlation probability fusion correction of all target centroids within the rectangular correlation gate, the equivalent centroid is obtained:

[0045]

[0046] in, This indicates the number of target centroids within the corrected rectangular associated gate.

[0047] In one embodiment of the present invention, existing tracks associated with the equivalent centroid are updated using the standard Kalman filter algorithm, and existing tracks not associated with the equivalent centroid are updated using the predicted values ​​in the standard Kalman filter algorithm, resulting in several updated tracks, including:

[0048] For existing tracks associated with the equivalent centroid, the prediction and update steps of the standard Kalman filter algorithm are used to update the existing tracks. For existing tracks not associated with the equivalent centroid, the prediction steps of the standard Kalman filter algorithm are used to update the existing tracks, resulting in several updated tracks.

[0049] In one embodiment of the present invention, the prediction step of the standard Kalman filter algorithm is as follows:

[0050]

[0051]

[0052] The update steps of the standard Kalman filter algorithm are as follows:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] in, To estimate the state values ​​a priori, To estimate the error covariance a priori, To estimate the state values ​​for the posterior time, To estimate the error covariance in the posterior timescale, To predict new information, To predict the covariance matrix of the new information, Here is the Kalman gain matrix. For the process noise covariance matrix and To observe the noise covariance matrix, For the current moment, For the previous moment, Here is the state transition matrix. For frame interval, This is the measurement matrix.

[0059] In one embodiment of the present invention, for target centroids not assigned to existing tracks, a logical method is used to initiate tracks, resulting in several stable initial tracks, including:

[0060] A rectangular associated gate is established with the target centroid point that has not been assigned to an existing track as the center point, and the equivalent centroid point of all target centroid points falling within the rectangular associated gate in the next frame is obtained. The center point and the equivalent centroid point are connected to establish several possible tracks.

[0061] A rectangular correlation gate is established centered on the extrapolated prediction point of each possible trajectory. When there is no target centroid point within the rectangular correlation gate, the possible trajectory is cancelled. When there is a target centroid point in the next frame within the rectangular correlation gate, the equivalent centroid points of all target centroid points are interconnected with the extrapolated prediction point, and this step is repeated until a stable starting trajectory is formed. The judgment logic of the stable starting trajectory is: a stable starting trajectory is formed when the ratio of the number of interconnected frames in the trajectory to the total number of frames is greater than or equal to a fourth threshold.

[0062] In one embodiment of the present invention, terminating the updated track or the stable initial track according to a termination condition includes:

[0063] When a preset number of frames are continuously predicted, the updated trajectory or the stable starting trajectory is terminated.

[0064] In one embodiment of the present invention, noise assessment is performed on the terminated trajectory to obtain the target tracking trajectory, including:

[0065] When the ratio of the predicted number of frames to the total number of tracking frames in the terminated track is greater than the fifth threshold, the terminated track is determined to be noise.

[0066] When the linear correlation coefficient of the terminated track in the radar line-of-sight direction is less than the sixth threshold, the terminated track is determined to be noise; the formula for calculating the linear correlation coefficient is:

[0067]

[0068] in, It is a time series. The mean of the time series. The coordinates are on the radar line of sight. This represents the average coordinates along the radar line of sight.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] 1. The present invention provides a millimeter-wave radar multi-vehicle tracking method based on traffic scenarios. This method utilizes a multi-level clustering algorithm based on target information for clustering, sets a rectangular association gate, and corrects the rectangular association gate using historical information obtained from target tracking. Finally, it uses a probabilistic data association algorithm to fuse the centroid points of the targets. Combining the characteristics of the radar measurement distribution and motion patterns of vehicle targets, it can accurately cluster neighboring targets when multiple vehicle targets are close to each other. It can effectively associate target tracks and measurements when there is a lot of scene clutter and a large number of targets, thus achieving robust tracking of multiple vehicle targets. This method solves the problem of difficult clustering and data association of densely adjacent targets that was not considered in the prior art, greatly improving the performance of multi-target clustering and data association, and enhancing the continuity and robustness of multi-vehicle target tracking in traffic scenarios.

[0071] 2. The method of this invention is based on target information and takes advantage of the longitudinal distribution of vehicle target measurements along the radar line of sight. It employs a multi-level clustering algorithm based on target information to cluster each frame of point cloud data and corrects the rectangular associated gate based on historical information obtained from target tracking. This achieves multi-dimensional hierarchical clustering, comprehensively utilizing multi-dimensional attribute information while preserving the physical meaning of clustering parameters. This ensures that all measurements in the clustering results come from the same target / clutter, effectively improving the clustering performance of adjacent vehicles and large vehicles in dense multi-vehicle situations. It solves the problem of inaccurate clustering results in current millimeter-wave radar multi-target tracking.

[0072] 3. The method of this invention combines target historical information with probabilistic data of the road environment. First, based on the road environment, a rectangular correlation gate is set according to the standard lane width and safe following distance, which reasonably and effectively transforms multi-target tracking into multiple single-target tracking. Then, based on the target historical information, the angle and position of the rectangular correlation gate are adjusted using the historical information obtained from target tracking. Finally, a probabilistic data correlation algorithm is used to fuse all target centroids within the corrected rectangular correlation gate according to probability, realizing effective correlation between target trajectory and measurement. Without increasing the algorithm complexity, it comprehensively utilizes the characteristics of the road environment and the movement law of vehicle targets along fixed routes, which greatly improves the accuracy of data correlation and solves the problem of poor data correlation performance in current millimeter-wave radar multi-target tracking, thereby realizing robust tracking of multiple vehicle targets in traffic scenarios. Attached Figure Description

[0073] Figure 1 A flowchart illustrating a millimeter-wave radar multi-vehicle tracking method based on a traffic scenario, provided as an embodiment of the present invention;

[0074] Figure 2 A flowchart illustrating another millimeter-wave radar multi-vehicle tracking method based on traffic scenarios provided in an embodiment of the present invention;

[0075] Figure 3 A framework diagram of a multi-level clustering algorithm based on target information provided in an embodiment of the present invention;

[0076] Figure 4 A schematic diagram of a probabilistic data association algorithm that combines historical target information and road environment, provided in an embodiment of the present invention;

[0077] Figure 5 This is a diagram of the actual data collection scenario for this invention. Detailed Implementation

[0078] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0079] Example 1

[0080] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating a millimeter-wave radar multi-vehicle tracking method based on a traffic scenario, provided by an embodiment of the present invention. Figure 2 This is a flowchart illustrating another millimeter-wave radar multi-vehicle tracking method based on a traffic scenario, provided as an embodiment of the present invention.

[0081] This embodiment of the millimeter-wave radar multi-vehicle tracking method in traffic scenarios includes the following steps:

[0082] S1. Preprocess the raw echo data from the millimeter-wave radar to obtain several frames of point cloud data.

[0083] Millimeter-wave radar transmits linear frequency modulated (LFM) continuous waves consisting of continuous LFM signals. The radar configuration parameters mainly include the number of transmitting antennas. Number of receiving antennas Carrier frequency ,bandwidth These parameters are determined by the distance resolution required in the actual measurement scenario. Speed ​​resolution Angular resolution Performance indicators were determined.

[0084] First, the raw echo data from the millimeter-wave radar is collected. Specifically, in traffic scenarios, the millimeter-wave radar is installed on overpasses or next to surveillance cameras, with its direction aligned with the road to monitor traffic conditions in both directions. Each transmitting antenna periodically emits a linear frequency modulated continuous wave, which is reflected by a moving vehicle and then... Each receiving antenna receives the echo signal, and each sample has a total of [number missing] antennas. Echo data from each channel.

[0085] Then, the raw echo data from the millimeter-wave radar is preprocessed to obtain scene-level point cloud data. The preprocessing specifically includes the following steps:

[0086] S11. The original echo data of the millimeter-wave radar is sequentially mixed, low-pass filtered, and sampled at discrete time to obtain a discrete-time signal.

[0087] Specifically, in each frame The echo signals from each channel are mixed with the corresponding transmitted signals using a quadrature mixer. The mixed signals are then low-pass filtered to obtain a complex exponential difference frequency signal. To facilitate subsequent digital signal processing, the difference frequency signal is sampled at discrete time to convert it into a discrete-time signal and saved.

[0088]

[0089]

[0090] in, For the first Difference frequency signal in discrete-time form of a single channel The amplitude of the difference frequency signal. The phase of the difference frequency signal. For discrete-time angular frequency, The number of discrete-time sampling points. Number of transmitting antennas Number of receiving antennas The frequency of the difference frequency signal. The discrete-time sampling frequency.

[0091] S12, regarding the obtained The discrete-time signals of each channel are sequentially ordered according to the number of sampling points in the fast time dimension. Number of sampling points in slow time dimension Number of antenna channels Arrange the signals to obtain a three-dimensional signal:

[0092]

[0093] S13. Perform discrete-time Fourier transforms on the three-dimensional signal sequentially according to the fast time dimension, slow time dimension, and antenna dimension to obtain the range-Doppler-angle (RDA) three-dimensional tensor:

[0094]

[0095] in, This represents the distance-Doppler-angle (RDA) tensor. This represents the discrete-time Fourier transform. This represents the number of points in the fast-time discrete-time Fourier transform (DFT). This represents the number of points in the slow-time discrete-time Fourier transform (DFT). This represents the number of discrete-time Fourier transform (DFT) points in the antenna channel dimension.

[0096] S14. Perform peak detection on the distance-Doppler-angle three-dimensional tensor. Obtain target status information:

[0097]

[0098] Here, CFAR stands for Constant False Alarm Rate (CFAR) detection method. Indicates the distance to the target. Indicates the Doppler velocity of the target. Indicates the azimuth angle of the target.

[0099] S15. Perform coordinate transformation on the target state information to obtain point cloud data for each frame of millimeter-wave radar. :

[0100]

[0101] in, The target point has two-dimensional coordinates. The velocity at the target point.

[0102] The millimeter-wave radar can be obtained through the above steps. Frame point cloud data .

[0103] S2. A multi-level clustering algorithm based on target information is used to cluster each frame of point cloud data, and the target centroid of each cluster is calculated.

[0104] With the first Taking a single frame point cloud as an example, due to factors such as low radar resolution, large target size, and different scattering intensities at different parts, the same target may be detected as multiple points in scene-level point cloud data. Therefore, it is necessary to obtain single frame point cloud data. Then, the point cloud data is clustered to obtain clusters of points representing multiple targets, thereby improving the accuracy of multi-target tracking in subsequent steps. Simultaneously, since tracking is performed on a single point, it is necessary to perform centroid aggregation on the clustered target point clusters to obtain the target centroid point. Centroids obtained by clustering frame point cloud data .

[0105] This embodiment addresses the clustering problem in multi-vehicle tracking under traffic conditions by designing a hierarchical clustering algorithm based on target information, building upon the traditional DBSCAN algorithm. Specifically, considering the longitudinal distribution of vehicle targets along the radar line of sight, rectangular region density is used instead of the circular region density of traditional DBSCAN for clustering, better describing the spatial attributes of vehicle targets. After obtaining the improved DBSCAN clustering results, the clustering results are adjusted (merged / split) by taking advantage of the fact that the measured speeds of different vehicle targets are basically consistent, thus achieving hierarchical clustering.

[0106] Please see Figure 3 , Figure 3 This is a framework diagram of a multi-level clustering algorithm based on target information provided in an embodiment of the present invention. Step S2 specifically includes the following steps:

[0107] S21. Input clustering parameters ,in, The parameters for target information are rectangular neighborhoods. The minimum number of points for clustering is determined by the type of target being tracked and the performance of the millimeter-wave radar.

[0108] Here, the rectangular neighborhood refers to the set of other samples centered at a sample point and whose distance to it does not exceed the rectangular area. That is:

[0109]

[0110] S22. Select any core sample point without a category from each frame of point cloud data, and find all sample sets that the core sample point can reach by density according to the clustering parameters to form a cluster. Repeat the clustering until all core sample points have a category to obtain several first point trace clusters.

[0111] First, select the scene level. Point cloud data Any core sample point without a category As a seed, find all the sample sets that the density of this core sample point can reach, which constitutes a cluster, and label the category of this cluster. Then, continue to select another core sample point without a category and find its density-reachable sample set to obtain another cluster and label its category. Repeat this step until all core sample points have a category, at which point the clustering is complete. A cluster of points is denoted as several first clusters of points.

[0112] Among them, the core sample point refers to any sample point If its rectangular neighborhood contains at least For each sample point, then These are called core sample points, and can be represented as:

[0113]

[0114] Density reachability means that if a point exists... , Depend on Density reaches directly, Depend on Density reaches directly, then and Density reachable. Density reachable means that if the sample points... If it is a cluster center, then its rectangular neighborhood sample point set All sample points from All are directly density-reachable. Density connectivity means that if points exist... , and They are all from If the density is achievable, then and Density connectivity.

[0115] Furthermore, a secondary clustering adjustment is performed based on the consistency statistics of the velocity vectors, splitting or merging the results of the primary clustering. This process includes steps S23 and S24.

[0116] S23. When the velocity variance of all points in the first point cluster is greater than the first threshold, the new clustering parameters are used to cluster all points in the first point cluster to obtain several second point clusters.

[0117] Specifically, for several first-point trace clusters, that is... For each cluster of traces, the variance of the velocity attributes of all traces within the cluster is calculated. When the variance exceeds a first threshold, the clustering result is considered unreasonable, potentially grouping multiple targets into one class. In this case, the first trace cluster of that class is re-clustered with smaller new clustering parameters, and the new clustering result is used as the splitting result for that class. A cluster of points is denoted as several second clusters of points.

[0118] S24. When the distance between the center points of two second point clusters is less than the second threshold and the velocity variance is less than the third threshold, the two second point clusters are merged to obtain several clusters.

[0119] Specifically, for several second-point trace clusters, i.e. The variance of the velocity attribute between each cluster of point traces is calculated. If two clusters are close enough and their variance is less than a certain threshold, then these two second point trace clusters are merged to obtain the final result. A cluster of points is denoted as several clusters.

[0120] It should be noted that steps S23 and S24 use velocity variance to reflect the consistency of velocity vectors, or the correlation coefficient or overlap coefficient of velocity can be used to reflect the consistency of velocity vectors.

[0121] S25. Perform centroid aggregation on each cluster, as follows:

[0122] Specifically, within the same cluster, when multiple core sample points exist, these core sample points are weighted according to their amplitude to obtain the target centroid. The amplitude of each point refers to its radar cross-section (RCS), meaning the RCS is used as the weight for centroid aggregation. When only one core sample point exists in the cluster, this core sample point is taken as the target centroid, and its coordinates are the coordinates of the aggregated target centroid. Centroids obtained by clustering frame point cloud data .

[0123] S3. Using the predicted point of the existing track as the center, and combining the standard lane width and safe following distance, set a rectangular associated gate. Based on the historical information obtained from the target tracking, correct the rectangular associated gate and predict the center position of the corrected rectangular associated gate.

[0124] After obtaining the target centroid, a data association process is required. This embodiment, targeting a traffic environment, sets up a rectangular association gate centered on the predicted point, taking into account the standard lane width and the safe following distance during driving. It then estimates the target heading angle using historical information obtained from target tracking, corrects the position and angle of the rectangular association gate using the heading information, and finally probabilistically weights the centroids within the gate to obtain the equivalent centroid for subsequent target tracking filtering.

[0125] Please see Figure 4 , Figure 4 A schematic diagram of a probabilistic data association algorithm that combines historical target information and road environment, provided in an embodiment of the present invention.

[0126] Step S3 specifically includes the following steps:

[0127] S31. Using the predicted point of the existing track as the center, set a rectangular associated gate in combination with the standard lane width and safe following distance.

[0128] Specifically, the first step is to analyze vehicle movement on a road. During their journey, vehicles generally travel along fixed lanes and maintain a safe braking distance from the vehicle in front to prevent accidents. Therefore, this prior information can be used to improve multi-vehicle target tracking performance. Considering these factors, this invention uses a rectangular correlation gate for data association, based on the standard lane width. safe following distance By setting the width and length of the rectangle and using appropriate rectangular correlation gates, the multi-target tracking problem in this scenario can be transformed into multiple single-target tracking problems.

[0129] S32. Correct the rectangular associated gate based on the historical information obtained from target tracking and predict the center position of the corrected rectangular associated gate.

[0130] The correlation gate set in the data association process is centered on the predicted point of the existing track. Therefore, whether the correlation gate can ensure that all measurement points of the target fall within it with a high probability depends not only on the shape and size of the gate, but also on the position coordinates of the predicted point and the target's heading angle. Therefore, this invention estimates the target's heading angle and corrects the predicted point position by tracking the target's historical information during the process, i.e., the velocity information in multiple frames of track data. Specifically, the steps include:

[0131] S321. For the existing track, calculate the current time before... The average velocity vector of the velocity vector in the frame track is used as the target heading angle, and the rectangular associated gate is rotated according to the target heading angle to obtain the corrected rectangular associated gate.

[0132] Specifically, for an existing track, take the current time. The former Velocity vector in frame track ,right The frame rate vector is averaged to obtain The direction of this vector is taken as the target heading angle, and the rectangular associated gate is rotated according to the target heading angle to make it conform to the target heading angle.

[0133] It should be noted that the above steps use the average velocity vector of the velocity vector as the target heading angle. Alternatively, the target heading angle can be determined by extracting the target's historical velocity at intervals or fitting the target's historical velocity from its historical position.

[0134] S322. Using the average velocity vector and the position information in the target state at the previous moment, predict the corrected rectangular associated gate to obtain the center position.

[0135] Specifically, when predicting existing tracks, the target state estimated at the previous moment is generally used for prediction. However, there is a problem that clutter can cause large estimation errors in the velocity component of the target state, leading to significant deviations in the predicted point position. Therefore, this embodiment obtains the average velocity vector... The velocity vector is used to correct the one-step prediction in the data association process. That is, the average velocity vector and the position information in the target state at the previous moment are used to make a one-step prediction, so as to obtain a more reliable center position of the association box as the center position of the positive rear rectangular association gate.

[0136] S4. Based on the center position, use a probabilistic data association algorithm to fuse all target centroids within the corrected rectangular associated gate according to probability to obtain an equivalent centroid.

[0137] Please combine Figure 4 For all centroids falling within the rectangular correlated gate The equivalent centroids are obtained by fusing data using a probabilistic data association algorithm based on probability. This is used for subsequent tracking filtering. The specific implementation process is as follows:

[0138] S41. Calculate the association probability of all target centroids within the corrected rectangular associated gate based on the center position. :

[0139]

[0140] in, This represents the probability of detecting the target. This represents the probability that the target appears within the corrected rectangular correlation gate. express The probability originating from the target, This indicates the centroid point that falls into the corrected rectangular correlated gate. Indicates the predicted center location. Describing covariance, This represents clutter density.

[0141] S42, Based on the aforementioned correlation probability The equivalent centroid is obtained by fusing and correcting all target centroids within the rectangular correlated gate. :

[0142]

[0143] in, This indicates the number of target centroids within the corrected rectangular associated gate.

[0144] S5. For existing tracks associated with the equivalent centroid, the standard Kalman filter algorithm is used to update the tracks. For existing tracks not associated with the equivalent centroid, the predicted values ​​in the standard Kalman filter algorithm are used to update the tracks, resulting in several updated tracks.

[0145] This embodiment uses the standard Kalman filter (KF) algorithm to update the track. The state equation and measurement equation of the standard Kalman filter algorithm are expressed as follows:

[0146]

[0147]

[0148] For vehicle targets in a road environment, this invention employs a uniform velocity model, with the state vector being... Simultaneously, by transforming radar measurements to a Cartesian coordinate system, the state transition matrix becomes... and measurement matrix for:

[0149] ,

[0150] in, The frame interval.

[0151] Furthermore, for existing tracks associated with the equivalent centroid, the prediction and update steps of the standard Kalman filter algorithm are used to update the existing tracks; for existing tracks not associated with the equivalent centroid, the prediction steps of the standard Kalman filter algorithm are used to update the existing tracks, resulting in several updated tracks.

[0152] The prediction and update steps of the standard Kalman filter (KF) algorithm are as follows:

[0153] predict:

[0154]

[0155]

[0156] renew:

[0157]

[0158]

[0159]

[0160]

[0161]

[0162] in, To estimate the state values ​​a priori, To estimate the error covariance a priori, To estimate the state values ​​for the posterior time, To estimate the error covariance in the posterior timescale, To predict new information, To predict the covariance matrix of the new information, Here is the Kalman gain matrix. For the process noise covariance matrix and To observe the noise covariance matrix, For the current moment, For the previous moment

[0163] S6. For target centroids not assigned to existing tracks, a logical method is used to initiate tracks, resulting in several stable initial tracks. Specifically, these include:

[0164] S61. Establish a rectangular associated gate with the target centroid point not assigned to an existing track as the center point, and obtain the equivalent centroid point of all target centroid points falling within the rectangular associated gate in the next frame. Connect the center point with the equivalent centroid point to establish several possible tracks.

[0165] S62. Establish a rectangular correlation gate centered on the extrapolated prediction point of each possible trajectory. When there is no target centroid point within the rectangular correlation gate, cancel the possible trajectory. When there is a target centroid point in the next frame within the rectangular correlation gate, obtain the equivalent centroid points of all target centroid points falling within the rectangular correlation gate in the next frame, connect the extrapolated prediction point with the equivalent centroid point, and repeat this step until a stable initial trajectory is formed. At this point, the trajectory initiation is complete. The logic for determining a stable initial trajectory is as follows: Logic: A stable initial track is formed when the ratio of interconnected frames to all frames in the track is greater than or equal to the fourth threshold. Within the time window of the frame The start of the trajectory is considered successful when the frames are interconnected. and The proportion is greater than the fourth threshold.

[0166] During the tracking process, centroids that do not fall into the rectangular associated gate are used as new track heads, and step S62 is restarted.

[0167] S7. Terminate the updated track or the stable starting track according to the termination condition, and perform noise judgment on the terminated track to obtain the target tracking track.

[0168] The multi-frame point cloud data obtained in step S1 is processed frame by frame, and steps S2-S6 are repeated frame by frame to obtain the tracking tracks of multiple vehicle targets. Track management is then performed on the tracking tracks of the multiple vehicle targets. Specifically, this includes:

[0169] S71, When continuously predicting a preset number of frames When this occurs, the updated track or the stable initial track is terminated. The preset number of frames... It is determined by the tracking scenario and the radar frame rate.

[0170] S72. The ratio of predicted frames to total tracking frames in the track after statistical termination. When this ratio is greater than the fifth threshold... When the flight path terminates, it is considered noise.

[0171] S73. Calculate the linear correlation coefficient of the terminated trajectory in the radar line-of-sight direction. When the linear correlation coefficient is less than the sixth threshold... When this happens, the terminated flight path is classified as noise. The formula for calculating the linear correlation coefficient is:

[0172]

[0173] in, It is a time series. The mean of the time series. The coordinates are on the radar line of sight. This represents the average coordinates along the radar line of sight.

[0174] Furthermore, this embodiment further illustrates the effectiveness of the present invention through experiments using measured data:

[0175] 1. Experimental conditions and experimental content

[0176] The software platform used in this experiment is: Windows 10 operating system, Matlab R2021b, Python 3.6, and PyTorch.

[0177] The hardware platform used in this experiment is: Dell T7910 workstation, CPU: Intel Core(TM) i7-4770, GPU: NVIDIA GTX 1080Ti.

[0178] The millimeter-wave radar parameters used in this experiment are: carrier frequency 80.05 GHz, bandwidth 73.728 MHz, range resolution 2.03 m, velocity resolution 0.043 m / s, angular resolution 0.8°, and maximum unambiguous velocity 78.95 km / h.

[0179] The experimental data used in this study were actual traffic scene data measured by millimeter-wave radar. The data was collected on a highway in a certain city. Figure 5 As shown, Figure 5 This is a scene diagram of the actual data collection for this invention. The scene is densely populated with vehicles, and the collection time is 200 seconds. Based on the synchronously collected video data, a total of 186 vehicle targets appeared during the collection period, and relevant tracking performance indicators were calculated based on the video data.

[0180] 2. Experimental Results and Analysis

[0181] Evaluation metrics for the test dataset:

[0182] Mean Track Lifetime (MTL) = Number of frames tracked on the target / Total number of frames in which the target appears (the average of all target track lifetimes is taken).

[0183] Tracking accuracy = Number of targets with a track lifetime greater than 80% / Total number of targets.

[0184] This embodiment aims to improve the performance of multi-vehicle target tracking in traffic scenarios. The experimental results are shown in Tables 1 and 2.

[0185] Table 1 Tracking performance under different methods

[0186]

[0187] As can be seen from Table 1, in the actual test of multi-vehicle target tracking on highways, the method proposed in this invention improves the average track lifetime and tracking accuracy by 15.3% and 28.3% respectively compared with the existing millimeter-wave radar multi-target tracking algorithm. This is because this invention proposes a millimeter-wave multi-vehicle tracking method based on traffic scenarios for multi-vehicle target tasks, which greatly improves the multi-target tracking performance in this scenario.

[0188] Table 2 Ablation experiments of the method proposed in this invention

[0189]

[0190] The method proposed in this invention addresses the clustering and association aspects of multi-target tracking algorithms for millimeter-wave radar and improves upon them for multi-vehicle tracking tasks in traffic scenarios. As shown in Table 2, the effectiveness of the two improvements is evident, both significantly enhancing multi-vehicle tracking performance in traffic scenarios.

[0191] This invention proposes for the first time a multi-vehicle tracking method based on traffic environment using millimeter-wave radar, which expands the application of millimeter-wave radar in real-world scenarios, improves tracking performance in traffic scenarios, and has high practical application value in the field of intelligent transportation.

[0192] This embodiment addresses the poor performance of tracking dense multi-vehicle targets in traffic environments by proposing a millimeter-wave radar multi-vehicle tracking method based on traffic scenarios. This method utilizes a multi-level clustering algorithm based on target information for clustering, sets a rectangular association gate, and corrects the rectangular association gate using historical information obtained from target tracking. Finally, it uses a probabilistic data association algorithm to fuse the target centroids. Combining the characteristics of radar measurement distribution and motion patterns of vehicle targets, it can accurately cluster neighboring targets when multiple vehicle targets are close to each other, and effectively associate target trajectories and measurements when there is a lot of scene clutter and a large number of targets, achieving robust tracking of multiple vehicle targets. This solves the problem of difficult clustering and data association of densely adjacent targets, which was not considered in existing technologies, greatly improving the performance of multi-target clustering and data association, and enhancing the continuity and robustness of multi-vehicle target tracking in traffic scenarios. Compared with existing technologies, it is more suitable for millimeter-wave radar multi-target tracking scenarios in the field of intelligent transportation, such as traffic flow monitoring and traffic accident early warning.

[0193] The method in this embodiment is based on target information and takes advantage of the longitudinal distribution of vehicle target measurements along the radar line of sight. It employs a multi-level clustering algorithm based on target information to cluster each frame of point cloud data and corrects the rectangular associated gate based on historical information obtained from target tracking. This achieves multi-dimensional hierarchical clustering, comprehensively utilizing multi-dimensional attribute information while preserving the physical meaning of clustering parameters. This ensures that all measurements in the clustering results come from the same target / clutter, effectively improving the clustering performance of adjacent vehicles and large vehicles in dense multi-vehicle situations. It solves the problem of inaccurate clustering results in current millimeter-wave radar multi-target tracking.

[0194] The method in this embodiment combines target historical information with probabilistic data of the road environment. First, based on the road environment, a rectangular correlation gate is set according to the standard lane width and safe following distance, which reasonably and effectively transforms multi-target tracking into multiple single-target tracking. Then, based on the target historical information, the angle and position of the rectangular correlation gate are adjusted using the historical information obtained from target tracking. Finally, a probabilistic data correlation algorithm is used to fuse all target centroids within the corrected rectangular correlation gate according to probability, realizing effective correlation between target trajectory and measurement. Without increasing the algorithm complexity, this method comprehensively utilizes the characteristics of the road environment and the movement patterns of vehicle targets along fixed routes, which greatly improves the accuracy of data correlation and solves the problem of poor data correlation performance in current millimeter-wave radar multi-target tracking, thereby achieving robust tracking of multiple vehicle targets in traffic scenarios.

[0195] In summary, the millimeter-wave radar multi-vehicle tracking method proposed in this embodiment, based on traffic scenarios, combines road environment, target measurement distribution characteristics, and target historical motion information. It designs a hierarchical clustering algorithm based on target information and a probabilistic data association algorithm combining historical information, improving the clustering performance of millimeter-wave radar for multi-vehicle targets and effectively associating target trajectories with measurements. This achieves robust and continuous tracking of multiple vehicle targets based on millimeter-wave radar, largely overcoming the bottlenecks and difficulties of multi-target tracking methods based on millimeter-wave radar in practical traffic scenarios. In the field of intelligent transportation, this method can perform functions such as real-time monitoring of vehicle flow and traffic conditions on roads, showing promising application prospects.

[0196] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A millimeter-wave radar multi-vehicle tracking method based on traffic scenarios, characterized in that, Including the following steps: The raw echo data from the millimeter-wave radar is preprocessed to obtain several frames of point cloud data. A hierarchical clustering algorithm based on target information is used to cluster each frame of point cloud data, and the target centroid of each cluster is calculated. Using the predicted point of the existing track as the center, a rectangular associated gate is set in combination with the standard lane width and safe following distance. The rectangular associated gate is corrected based on the historical information obtained from target tracking, and the center position of the corrected rectangular associated gate is predicted. Based on the aforementioned center position, a probabilistic data association algorithm is used to fuse all target centroids within the corrected rectangular associated gate according to probability, thereby obtaining an equivalent centroid. For existing tracks associated with the equivalent centroid, the prediction and update steps of the standard Kalman filter algorithm are used to update the existing tracks. For existing tracks not associated with the equivalent centroid, the prediction values ​​in the standard Kalman filter algorithm are used to update the existing tracks, resulting in several updated tracks. For target centroids that have not been assigned to existing tracks, a logical method is used to initiate the track, resulting in several stable initial tracks. The updated track or the stable starting track is terminated according to the termination condition, and noise is judged on the terminated track to obtain the target tracking track.

2. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, Preprocessing the raw echo data from the millimeter-wave radar yields several frames of point cloud data, including: The raw echo data from the millimeter-wave radar is sequentially mixed, low-pass filtered, and sampled at discrete time to obtain a discrete-time signal: in, For the first Difference frequency signal in discrete-time form of a single channel The amplitude of the difference frequency signal. The phase of the difference frequency signal. For discrete-time angular frequency, The number of discrete-time sampling points. Number of transmitting antennas Number of receiving antennas The frequency of the difference frequency signal. The discrete-time sampling frequency; The discrete-time signals are arranged sequentially according to the number of sampling points in the fast time dimension, the number of sampling points in the slow time dimension, and the number of antenna channels to obtain a three-dimensional signal: in, For the number of sampling points in the fast time dimension, The number of sampling points in the slow time dimension. This refers to the number of antenna channels; Perform discrete-time Fourier transforms on the three-dimensional signal sequentially according to the fast time dimension, slow time dimension, and antenna dimension to obtain the range-Doppler-angle three-dimensional tensor: in, This represents the distance-Doppler-angle tensor. This represents the discrete-time Fourier transform. This represents the number of points in the fast-time discrete-time Fourier transform. This represents the number of points in the slow-time discrete-time Fourier transform. The number of discrete-time Fourier transform points representing the antenna channel dimension; Peak detection is performed on the distance-Doppler-angle three-dimensional tensor to obtain target state information: Here, CFAR stands for Constant False Alarm Rate (CFAR) detection method. Indicates the distance to the target. Indicates the Doppler velocity of the target. Indicates the target's azimuth angle; The target state information is transformed to obtain point cloud data for each frame: in, The target point has two-dimensional coordinates. The velocity at the target point.

3. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, A hierarchical clustering algorithm based on target information is used to cluster each frame of point cloud data, and the target centroid of each cluster is calculated, including: Input clustering parameters ,in, Parameters for target information, To minimize the number of points in the cluster, the target information is expressed as a rectangular neighborhood; Select any core sample point without a category from each frame of point cloud data, and find all sample sets that the core sample point can reach in density according to the clustering parameters to form a cluster. Repeat the clustering until all core sample points have a category to obtain several first point trace clusters. When the velocity variance of all points in the first point cluster is greater than the first threshold, the new clustering parameters are used to cluster all points in the first point cluster to obtain several second point clusters. When the distance between the center points of two second point trace clusters is less than the second threshold and the velocity variance is less than the third threshold, the two second point trace clusters are merged to obtain several clusters. When there are multiple core sample points in the cluster, the multiple core sample points are weighted according to the magnitude of each point to obtain the target centroid point; when there is only one core sample point in the cluster, the core sample point is used as the target centroid point.

4. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, Based on historical information obtained from target tracking, the rectangular correlated gate is corrected and the center position of the corrected rectangular correlated gate is predicted, including: For the existing track, calculate the current time before... The average velocity vector of the velocity vector in the frame track is used as the target heading angle, and the rectangular associated gate is rotated according to the target heading angle to obtain the corrected rectangular associated gate; The center position is obtained by predicting the corrected rectangular associated gate using the average velocity vector and the position information of the target state at the previous moment.

5. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, Based on the aforementioned center position, a probabilistic data association algorithm is used to fuse all target centroids within the corrected rectangular associated gate according to probability, resulting in equivalent centroids, including: Calculate the association probability of all target centroids within the corrected rectangular associated gate based on the center position: in, This represents the probability of detecting the target. This represents the probability that the target appears within the corrected rectangular correlation gate. express The probability originating from the target, This indicates the centroid point that falls into the corrected rectangular correlated gate. Indicates the predicted center location. express covariance, Indicates clutter density, Indicates the number of sampling points in the fast time dimension; Based on the correlation probability fusion correction of all target centroids within the rectangular correlation gate, the equivalent centroid is obtained: in, This indicates the number of target centroids within the corrected rectangular associated gate.

6. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, The prediction steps of the standard Kalman filter algorithm are as follows: The update steps of the standard Kalman filter algorithm are as follows: in, To estimate the state values ​​a priori, To estimate the error covariance a priori, To estimate the state values ​​for the posterior time, To estimate the error covariance in the posterior timescale, To predict new information, To predict the covariance matrix of the new information, Here is the Kalman gain matrix. For the process noise covariance matrix and To observe the noise covariance matrix, For the current moment, For the previous moment, Here is the state transition matrix. For frame interval, This is the measurement matrix.

7. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, For target centroids not assigned to existing tracks, a logical method is used to initiate tracks, resulting in several stable initial tracks, including: A rectangular associated gate is established with the target centroid point that has not been assigned to an existing track as the center point, and the equivalent centroid point of all target centroid points falling within the rectangular associated gate in the next frame is obtained. The center point and the equivalent centroid point are connected to establish several possible tracks. A rectangular correlation gate is established centered on the extrapolated prediction point of each possible trajectory. When there is no target centroid point within the rectangular correlation gate, the possible trajectory is cancelled. When there is a target centroid point in the next frame within the rectangular correlation gate, the equivalent centroid points of all target centroid points are interconnected with the extrapolated prediction point, and this step is repeated until a stable starting trajectory is formed. The judgment logic of the stable starting trajectory is: a stable starting trajectory is formed when the ratio of the number of interconnected frames in the trajectory to the total number of frames is greater than or equal to a fourth threshold.

8. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, Terminate the updated track or the stable starting track according to the termination conditions, including: When a preset number of frames are continuously predicted, the updated trajectory or the stable starting trajectory is terminated.

9. The millimeter-wave radar multi-vehicle tracking method based on traffic scenarios according to claim 1, characterized in that, Noise assessment is performed on the terminated trajectory to obtain the target tracking trajectory, including: When the ratio of the predicted number of frames to the total number of tracking frames in the terminated track is greater than the fifth threshold, the terminated track is determined to be noise. When the linear correlation coefficient of the terminated track in the radar line-of-sight direction is less than the sixth threshold, the terminated track is determined to be noise; the formula for calculating the linear correlation coefficient is: in, It is a time series. The mean of the time series. The coordinates are on the radar line of sight. This represents the average coordinates along the radar line of sight.

Images