Millimeter wave radar tracking method in multi-target complex environment

Abstract

The invention belongs to the technical field of radar signal back-end processing, and relates to a millimeter wave radar tracking method in a multi-target complex environment, which comprises the following steps of: 1, processing a radio frequency signal received by a radar to generate millimeter wave radar point cloud data; 2, delimiting a radar action range, removing point cloud data beyond the range, and then carrying out point cloud clustering through a DBSCAN clustering method to obtain merge points; and step 3, starting a track, matching measured data with the track through data association gate judgment, if association succeeds, updating a target state and then outputting the track, and otherwise, executing track loss or maintenance operation by adopting a scoreboard mechanism. According to the method, the accuracy of target tracking can be effectively improved, the error tracking rate of multi-target tracking can be reduced, meanwhile, the calculation complexity of an algorithm is reduced, and the calculation time is greatly shortened.

Description

Technical Field

[0001] The invention belongs to the technical field of radar signal back-end processing and relates to a millimeter wave radar tracking method in a multi-target complex environment. Background Art

[0002] As the information level of today's society becomes higher and higher, the application of millimeter-wave radar technology is mainly divided into two aspects: military and civilian. The civilian field is often used for signal processing, such as measuring the distance, speed, azimuth and other information of the target, and performing back-end processing such as clustering and tracking through the generated radar point cloud data after signal processing. There are also functions such as gesture recognition, anti-fall detection and heartbeat detection through micro-Doppler features. These are inseparable from its core device, the millimeter-wave radar sensor. Therefore, the radar industry has developed rapidly and shined in the fields of military, transportation, automobiles, medical care, elderly care, drones, etc.

[0003] Radar's multi-target tracking is mainly used in scenarios such as traffic and security to detect and track unknown people or vehicles. Due to the strong anti-interference ability of millimeter-wave radar, it can more accurately identify moving targets at night or in harsh environmental conditions.

[0004] Since the target is tracked in a dense clutter environment, the most important issue is the accuracy of single target tracking and the problem of mistracking when tracking multiple targets. When two or more objects are too close to each other, such as in parallel motion or intersecting motion, using common data association algorithms, such as associating the measurement data of the current frame with the trajectory of the previous frame, often results in mistracking. In the prior art, distance-based data association algorithms are often used, such as the NNDA nearest neighbor association algorithm and the GNNDA global nearest neighbor algorithm. This method takes the closest observation point as the real target point, which will result in a target point being determined to be associated with two different trajectories when two targets are close, resulting in mistracking. Summary of the invention

[0005] In order to solve the above technical problems existing in the prior art, the present invention proposes a millimeter wave radar tracking method in a multi-target complex environment, and its specific technical solution is as follows:

[0006] A millimeter wave radar tracking method in a multi-target complex environment includes the following steps:

[0007] Step 1: Process the radio frequency signal received by the radar to generate millimeter wave radar point cloud data;

[0008] Step 2: Define the radar range, remove the point cloud data beyond the range, and then cluster the point cloud using the DBSCAN clustering method to obtain the merge point;

[0009] Step 3: Perform track initiation. Match the measurement data with the track by judging through the data association gate. If the association is successful, update the target state and output the track. Otherwise, execute the operations of track loss or maintenance using the scoreboard mechanism.

[0010] Further, Step 1 specifically includes the following sub-steps:

[0011] Step 1.1: Separate the data of different channels in the radio frequency signal, and improve the signal-to-noise ratio SNR of the signal by removing the DC component, filtering and noise reduction, and matched filtering.

[0012] Step 1.2: Perform a fast Fourier transform on the signal in the distance dimension, and process the echo signal in the time domain through pulse compression technology to calculate the distance information of the target.

[0013] Step 1.3: Perform Doppler FFT processing on the slow-time data to extract the Doppler frequency of the target, so as to obtain the radial velocity of the target.

[0014] Step 1.4: Remove static targets through moving target indication MTI filtering technology.

[0015] Step 1.5: Perform two-dimensional constant false alarm rate detection CFAR to screen out the target points and their intensity information.

[0016] Step 1.6: Perform direction of arrival DOA estimation to obtain the target angle information.

[0017] Further, in Step 2, when delimiting the radar coverage area and removing the point cloud data outside the range, specifically: set the starting origin, angle measurement range, and distance measurement range, and judge whether it is within the range according to the calculated Euclidean distance between the measurement point and the starting origin and the angle, so as to obtain the processed point cloud.

[0018] Further, in Step 3, the m / n logic method is adopted for track initiation, specifically: obtain all the measurements at the initial moment, that is, the measured values, and establish a circular gate of the maximum distance and the minimum distance according to the estimated maximum speed, minimum speed, and sampling interval;

[0019] Associate the second-frame measurement falling within the initial correlation gate with the initial measurement to form a candidate track;

[0020] Extrapolate the candidate track formed at the previous moment according to the Kalman filter KF algorithm. Taking the extrapolated point as the center, confirm the size of the subsequent gate according to the track extrapolation error covariance, so as to form the sampling confirmation area at the next moment. If the third-frame measurement falls within the correlation gate, perform the association. If multiple measurements fall within it, take the measurement closest to the extrapolated point;

[0021] Continue to extrapolate. If it is found that no measurement falls within the gate of the extrapolated point, record the number of failures, and use the extrapolated point as the predicted value of the current frame for extrapolation. If the number of failures is greater than n - m times, it is proved to be a failure. Continue to extrapolate until a stable track is formed and the track initiation is successful.

[0022] Further, the Kalman filter KF algorithm specifically includes:

[0023] Represent the target state vector of the entire multi-target tracking process as:

[0024] X k = [x, y, x', y']

[0025] Where x and y represent the position information of the target, and x' and y' represent the velocity information of the target;

[0026] Prediction step: Predict the state at the next moment based on the final predicted value at the previous moment;

[0027] State transition equation: X k|k-1 = FX k-1|k-1 + w k

[0028] Predicted covariance matrix: P k|k-1 = FP k-1|k-1 F T + Q

[0029] In the formula, X k-1|k-1 represents the final predicted value at the previous moment, w k is the process noise, Q is the process noise covariance matrix, and F is the state transition matrix;

[0030] Update step: Adjust the predicted value of the prediction step according to the measurement value to obtain the final state prediction value;

[0031] Measurement equation: Z k = HX k + u k

[0032] Measurement covariance matrix: S k = HP k|k-1 H T + R

[0033] Kalman gain:

[0034] State update: X k|k = X k|k-1 + K k (Z k - HX k|k-1 )

[0035] Updated covariance matrix: Pk|k =(I - K k H)P k|k-1

[0036] where R is the measurement noise covariance matrix, Z k - HX k|k-1 is the information at this moment, and the measurement matrix H is:

[0037]

[0038] Take the currently updated state X k|k as the input for the next moment, and repeat the prediction and update processes to achieve radar target tracking.

[0039] Furthermore, in step three, the matching of measurement data with the track by judging through the data association gate is implemented using the GJPDA algorithm. The specific steps of this algorithm are as follows:

[0040] Step 1: Use the Mahalanobis distance as the criterion, make a prediction based on the estimated value of the previous moment, and establish a gate centered on the predicted value to obtain the effective measurements falling within the gate

[0041] Step 2: Obtain the confirmation matrix Ω from the relationship between the effective measurements and the target confirmation gate;

[0042] Step 3: Calculate the association probability β between the target and the measurements according to the likelihood function jt ;

[0043] Step 4: Obtain the final combined information v k , and according to the Kalman filter update step, the target state estimate value and covariance matrix at the current moment can be obtained and wait to be used in the next iteration.

[0044] Furthermore, in step 2, the establishment process of the confirmation matrix Ω is as follows:

[0045]

[0046] where w jt represents the flag indicating whether the measurement j (j = 1, 2,..., m k ) is associated with the target t (t = 0, 1, 2,..., T), and its value is 0 or 1; when w jt = 0, it means that the j - th measurement is not associated with the t - th target; when w jt = 1, it means that the j - th measurement is associated with the t - th target; the first column of the confirmation matrix is all 1, indicating that at a certain moment, no target is associated with the measurement, and all measurements are false measurements; m kis the number of measurements falling within the tracking gate at time k, and T is the number of targets.

[0047] Furthermore, in step 3, the likelihood function is used instead of establishing multiple association matrices and calculating the conditional probability of each association matrix to obtain the association probability between measurement j and target t. The calculation formula for the final association probability in this algorithm is as follows;

[0048]

[0049] In the formula, G jt can be regarded as the probability that the j-th measurement is associated with the t-th target;

[0050]

[0051] In the formula represents the information function, which is used to calculate the information between the j-th measurement and the t-th target; S jt (k) represents the information covariance corresponding to the information function; in the actual tracking situation, the value of B is related to the clutter density of the environment.

[0052] Furthermore, in step 4, the combined information is calculated as follows:

[0053]

[0054] According to the Kalman filter, calculate the final state estimate value of target t at time k and update the posterior covariance.

[0055] Z k|k = X k|k-1 + K k V t

[0056] P k∣k = P k∣k-1 -(1 - β 0t )K k S k K k T + K k (Z k|k - V t V t T )K k T

[0057] If the track in step 1 is a failed track, then X k|k-1 is used as the final state estimate value and input into the next loop, and the covariance matrix at the previous moment is used as the covariance matrix P k∣k .

[0058] Further, the scoreboard mechanism is specifically as follows: record the number of times the target is lost and decide whether to continue maintaining the track; if the number of times the target is lost exceeds the set threshold n, it is considered that the target has been lost, perform the track loss operation, and terminate the tracking of the track; otherwise, enter the track maintenance stage, continue to wait for subsequent measurement data to determine whether the target still exists.

[0059] Beneficial effects:

[0060] The edge track initiation in the present invention can effectively solve the problem of generating duplicate tracks similar to the true target track due to excessive clutter density or too small a track gate. By using the common sense that new targets only appear at the edge of the radar's operating range, the computational complexity of the algorithm can be effectively reduced.

[0061] The GJPDA proposed in the present invention, based on the commonly used JPDA algorithm for multi-target tracking, optimizes the step of establishing the association matrix, and uses the likelihood function to replace the calculation of the conditional probability of the original association matrix, effectively solving the situation where the number of observations increases due to the increase in clutter density, resulting in an exponential increase in the association matrix, greatly reducing the calculation time. At the same time, the step of the confirmation matrix is retained, well preserving the multi-target tracking characteristics of JPDA. If deployed on an embedded device, it can also solve the real-time problem to a certain extent.

[0062] The present invention can effectively match new measurement data with the correct target track through gate judgment and m / n association rules, reducing the possibility of incorrect association. The scoreboard mechanism is used to evaluate the stability of the track, preventing premature termination of target tracking due to short-term measurement loss. Description of the Drawings

[0063] Figure 1 is the radar multi-target tracking flowchart of the embodiment of the present invention;

[0064] Figure 2 is the radar signal processing flowchart of the embodiment of the present invention;

[0065] Figure 3 is the schematic diagram of the DBSCAN algorithm of the embodiment of the present invention;

[0066] Figure 4 is the track management flowchart of the embodiment of the present invention;

[0067] Figure 5 is the schematic diagram of the m / n logical track initiation of the embodiment of the present invention;

[0068] Figure 6 is the schematic diagram of the confirmation matrix establishment process of the GJPDA algorithm of the embodiment of the present invention;

[0069] Figure 7 It is a collection diagram of the pedestrian cross - movement dataset in the embodiment of the present invention;

[0070] Figure 8 It is the display of the radar point cloud data of the pedestrian cross - movement in the embodiment of the present invention;

[0071] Figure 9 It is the multi - target tracking result diagram of the pedestrian cross - movement in the embodiment of the present invention;

[0072] Figure 10 It is a collection diagram of the bicycle close - range parallel - movement dataset in the embodiment of the present invention;

[0073] Figure 11 It is the display of the radar point cloud data of the bicycle close - range parallel - movement in the embodiment of the present invention;

[0074] Figure 12 It is the multi - target tracking result diagram of the bicycle close - range parallel - movement in the embodiment of the present invention. Specific implementation manner

[0075] In order to make the purpose, technical solution and technical effect of the present invention clearer and more understandable, the following further elaborates on the present invention in combination with the specification drawings and embodiments.

[0076] As Figure 1 shown, this embodiment discloses a millimeter - wave radar tracking method in a multi - target complex environment, including the following steps:

[0077] Step 1: Perform signal processing on the radio - frequency data received by the radar to generate millimeter - wave radar point cloud data.

[0078] As Figure 2 shown, first, in the data pre - processing stage, the MIMO radar, through the configuration of N transmitting antennas and M receiving antennas, can approximately simulate N×M virtual channels. The formation of such virtual channels significantly improves the spatial resolution and target detection ability of the radar. In this stage, the main task of data pre - processing is to separate the data of different channels in the radio - frequency signal, and improve the signal - to - noise ratio SNR of the signal through methods such as removing the DC component, filtering and noise reduction, and matched filtering, providing a better data basis for subsequent processing.

[0079] After completing the data preprocessing, the signal enters the range-Doppler processing stage to achieve incoherent accumulation. In this stage, the signal is first subjected to a fast Fourier transform FFT in the distance dimension, and the echo signal is processed in the time domain through pulse compression technology to accurately calculate the distance information of the target. Next, the slow-time data is subjected to Doppler FFT processing to extract the Doppler frequency of the target, thereby obtaining the radial velocity of the target. At the same time, in order to eliminate interference signals in static scenes, such as echoes from the ground or fixed objects, the radar system will use the moving target indication MTI (Moving Target Indication) filtering technology to remove static targets and make the signals of dynamic targets more prominent. This dynamic filtering process not only improves the system's sensitivity to moving targets, but also significantly reduces the impact of background noise.

[0080] After MTI filtering, the static clutter in the signal has been effectively filtered out, and the signal will enter the two-dimensional constant false alarm rate (CFAR) detection stage. The core idea of ​​the CFAR detection algorithm is to maintain a constant false alarm probability during the target detection process, and to ensure the stability of the system in different environments by dynamically adjusting the detection threshold. Commonly used two-dimensional CFAR algorithms include unit average CFAR (CA-CFAR) and maximum selection CFAR (OS-CFAR). These algorithms can adaptively identify the location of the target point in the range-Doppler two-dimensional data. During the detection process, the CFAR algorithm constructs a sliding window to perform statistical analysis on the noise environment around the target, and calculates the detection threshold based on the set false alarm probability, and finally screens out possible target points and their intensity information.

[0081] After completing the detection of the target point, in order to further accurately obtain the spatial position information of the target, the radar system will enter the DOA (Direction of Arrival) estimation stage. In this stage, the commonly used spatial spectrum estimation MUSIC (Multiple Signal Classification) algorithm is based on the separation of signal subspace and noise subspace. By constructing a spatial spectrum function and scanning at the spatial angle, the angle information of the target is inferred using the maximum value method. The MUSIC algorithm can provide extremely high angular resolution under high signal-to-noise ratio conditions, and is particularly suitable for systems such as MIMO radars with a large number of receiving channels.

[0082] After the above steps, more accurate millimeter-wave radar point cloud information can be obtained.

[0083] Step 2: Define the radar range, remove the observations that are out of range, and then cluster the point cloud using the DBSCAN clustering method to cluster the ideal merge points.

[0084] Specifically, a preprocessing module is set up to read the point cloud information obtained from radar signal processing, cluster the point cloud, and remove the average value as the ideal merge point. For the clustering method, the DBSCAN clustering method based on density clustering is adopted, and the K-nearest neighbor is used to obtain the K-th nearest distance as the radius in DBSCAN to achieve adaptive clustering. At the same time, the radar range is automatically divided, and when new measurement values are input, the observed values exceeding the radar range are directly excluded. Additionally, in the real situation, after the track head starts, if new targets or tracks appear later, they generally enter from the edge of the radar detection range. Therefore, an edge track start mechanism is proposed to reuse the data preprocessing module and overload the function that proposes values exceeding the radar measurement range.

[0085] For example, set the starting origin {0, 0}, the angle measurement range {10, 150}, and the ranging range {0, 50}. According to the calculation of the Euclidean distance between the measured point and the starting origin, and the angle to determine whether it is within the range, the processed point cloud is obtained. The Euclidean distance formula is as follows:

[0086]

[0087] After that, a clustering operation is performed. Here, the DBSCAN algorithm is used, and this algorithm is determined by the following two parameters:

[0088] ε (eps): Defines the size range of the neighborhood;

[0089] MinPts: The minimum number of neighborhood points for a point to be considered a core point.

[0090] The DBSCAN algorithm is density-based clustering. If p is a core point, then there are at least MinPts points within the range of ε. If q is also a core point and q is within the range of ε of p, then it is said that q is directly density-reachable from p, and so on to form a cluster composed of multiple core points and neighborhood points, as Figure 3 shown, and unlike the clusters formed by ordinary Euclidean clustering which are regular circles, this has better recognition for objects that do not belong to regular shapes, and does not require knowing the number of targets in advance, and automatically identifies into clusters with different numbers.

[0091] Then, according to the averages of all x and y of these clusters, the ideal merge point is obtained.

[0092] Step three: As Figure 4As shown in the figure, based on the information of the existing track and the measurement input, i.e., the new measurement data provided by the sensor, it is determined whether the current measurement value is within the track gate. The track gate is a threshold area generated according to the predicted target state, and only the measurement values falling within this area may be associated with the track. If the measurement value is within the track gate, the filtering prediction step is executed, and methods such as Kalman filtering are used to update the target state to improve the accuracy of subsequent data association. Otherwise, it is further checked whether the current measurement value meets the m / n logic association condition. If this condition is satisfied, new track initiation is performed, i.e., a new target track is created.

[0093] Furthermore, data association gate judgment is carried out, that is, it is checked whether the measurement value is within a more stringent data association gate range to determine whether it can be matched with the existing track. If the measurement value is within the data association gate, data association is performed, that is, the measurement data is matched with the existing track to update the target state. After successful association, the filtering update step is executed to correct the target state and improve the accuracy of track estimation. Finally, track output is performed to provide stable target trajectory information.

[0094] If the measurement value fails to be associated with any existing track, a scoreboard mechanism is adopted for management to continuously evaluate the stability and reliability of the track; this mechanism is used to record the number of times the target is lost and decide whether to continue maintaining the track. If the number of times the target is lost exceeds the set threshold n, it is considered that the target has been lost, and the track loss operation is executed to terminate the tracking of the track; otherwise, it enters the track maintenance stage and continues to wait for subsequent measurement data to determine whether the target still exists. The whole process is executed in a loop, which can stably track the target in a noisy environment and is applicable to various application scenarios such as radar and visual target tracking.

[0095] Based on the above method, in this embodiment, a track management module is correspondingly designed, as well as a tracking filtering module and a data association module that run in parallel therein.

[0096] As Figure 5 shown, the track initiation, i.e., the process of generating a track header, in this embodiment uses m / n logic track initiation, that is, within n frames, if the same target is observed at least m times, it is considered that the target exists, and an initial track is generated in the order of m frames, and the covariance matrix of the last frame is obtained by relying on Kalman filtering.

[0097] The specific process of the m / n logic track initiation is as follows:

[0098] Obtain all the measurements, i.e., measurement values, at the initial moment, and establish the maximum distance V max and the minimum distance V min according to the estimated maximum speed V max *T and the minimum speed V min*Ring gate of T:

[0099] d ij (t) = max[0, z j (k + 1) - z i (k) - v max t] + max[0, -z j (k + 1) + z i (k) + v min t]

[0100]

[0101] Associate the second-frame measurement falling within the gate with the initial measurement to form a candidate track.

[0102] Extrapolate the candidate track formed at the previous moment according to the Kalman filter KF algorithm. Taking the extrapolated point as the center, confirm the size of the subsequent gate based on the track extrapolation error covariance, thereby forming the sampling confirmation area at the next moment. If the third-frame measurement falls within the relevant gate, perform the association. If multiple measurements fall within, select the measurement closest to the extrapolated point.

[0103] Continue extrapolating. If it is found that no measurement falls within the gate of the extrapolated point, record the failure count, and use the extrapolated point as the predicted value of the current frame for extrapolation. If the failure count is greater than n - m times, it is proved to be a failure, and continue extrapolating until a stable track is formed and the track initiation is successful.

[0104] Among them, in the filtering prediction step executed by the tracking filtering module, the target position and covariance matrix at the previous moment are used to predict the position where the target should be at the current moment. In this embodiment, the Kalman filter (KF) is used as the main algorithm to achieve target tracking. First, establish a state model, including state variables such as position and velocity, and define the state transition matrix to predict the position of the target at the next moment. Subsequently, calculate the prediction uncertainty P using the process noise covariance matrix. When the sensor obtains the measurement value, calculate the measurement residual through the measurement matrix H. Then, calculate the Kalman gain for weighted fusion of the predicted value and the measurement value, update the state estimate, and adjust the uncertainty. This process loops continuously to make the target trajectory estimate more accurate. The implementation process of the KF algorithm specifically includes:

[0105] Represent the target state vector of the entire multi-target tracking process as:

[0106] X k = [x, y, x', y']

[0107] Where x and y represent the position information of the target, and x' and y' represent the velocity information of the target.

[0108] Prediction step: Predict the state at the next moment based on the final predicted value at the previous moment.

[0109] State transition equation: X k|k-1 = FX k-1|k-1 + w k

[0110] Predicted covariance matrix: P k|k-1 = FP k-1|k-1 F T + Q

[0111] Wherein, X k-1|k-1 represents the final predicted value at the previous moment, w k is the process noise, Q is the process noise covariance matrix, F is the state transition matrix, and here it can be obtained according to the target state vector:

[0112]

[0113] where T is the sampling interval of the radar.

[0114] Update step: Adjust the predicted value in the prediction step according to the measurement value to obtain the final state predicted value.

[0115] Measurement equation: Z k = HX k + u k

[0116] Measurement covariance matrix: S k = HP k|k-1 H T + R

[0117] Kalman gain:

[0118] State update: X k|k = X k|k-1 + K k (Z k - HX k|k-1 )

[0119] Update covariance matrix: P k|k = (I - K k H)P k|k-1

[0120] Wherein, R is the measurement noise covariance matrix, Z k - HX k|k-1 is the information at this moment, and the measurement matrix H is:

[0121]

[0122] Take the currently updated state X k|k as the input for the next moment, and repeat the prediction and update process to achieve radar target tracking.

[0123] In the data association module, the measured values at the current moment are associated with the predicted values, and multiple observed values are accurately matched with multiple target predicted values in a multi-target environment. For multiple objects running in parallel or intersecting, it is often impossible to accurately determine only by distance. Therefore, a data association algorithm JPDA (Joint Probability Density Association) specifically used for multi-target tracking is selected. This algorithm is an advanced version of PDA (Probability Density Association). PDA is only applicable to the single-target case. Therefore, on this basis, by generating a confirmation matrix with m rows and n columns, where there are m targets and n measurements, multiple interconnection matrices can be generated according to certain rules. By calculating the probability of each interconnection matrix, it is finally determined which measurement ultimately belongs to which target. However, the drawback of this algorithm is that as the number of targets and measurements increases, the number of interconnection matrices will increase exponentially. If deployed on hardware devices, it cannot meet the real-time requirements. Therefore, in this embodiment, a method for approximately calculating the interconnection probability through the likelihood function is proposed to reduce the number of considered association cases, thereby improving the efficiency of the algorithm. This algorithm is called GJPDA. When calculating the association probability, the GJPDA algorithm has the characteristics of the JPDA algorithm. For measurements that only appear within the tracking gate of one target, a higher weight will be given, while for measurements within the overlapping region of multiple target tracking gates, their weights will be reduced. This weight assignment strategy can effectively handle the overlapping situation of target-related gates and improve the robustness of the algorithm. The implementation process of the GJPDA algorithm includes the following steps:

[0124] Step 1: Predict based on the estimated value at the previous moment, and establish a gate with the center of the predicted value to obtain the valid measurements that fall within the gate

[0125] Step 2: As shown in Figure 6 , obtain the confirmation matrix Ω from the relationship between the valid measurements and the target confirmation gate

[0126] Step 3: Calculate the interconnection probability β between the target and the measurement according to the likelihood function jt .

[0127] Step 4: Obtain the final combined information v k , and according to the Kalman filter update step, the target state estimate value and covariance matrix at the current moment can be obtained, waiting to be used in the next iteration

[0128] The formula for establishing the gate with the center of the predicted value in Step 1 is as follows. Here, the Mahalanobis distance is used as the determination. The predicted value at the previous moment is used as the center of the gate, and γ is used as the radius

[0129]

[0130] In the formula, S-1 (k) is the covariance matrix of the target prediction at the previous moment, is the predicted value at this moment obtained based on the target at the previous moment, z i (k) is the observed value at this moment.

[0131] In step 2, as Figure 6 shown, the establishment process of the confirmation matrix Ω is as follows:

[0132]

[0133] where w jt represents the flag indicating whether the measurement j (j = 1, 2,..., m k ) is associated with the target t (t = 0, 1, 2,..., T), and its value is 0 or 1. Specifically, when w jt = 0, it means that the j-th measurement is not associated with the t-th target; when w jt = 1, it means that the j-th measurement is associated with the t-th target. The first column of the confirmation matrix is all 1, indicating that at a certain moment, no target is associated with the measurement, and all measurements are false measurements. m k is the number of measurements falling within the tracking gate at time k, and T is the number of targets.

[0134] In step 3, the likelihood function is used instead of the method of establishing multiple association matrices and calculating the conditional probability of each association matrix to obtain the association probability between the measurement j and the target t. The calculation formula of the final association probability in this algorithm is as follows;

[0135]

[0136] In the formula, G jt can be regarded as the probability that the j-th measurement is associated with the t-th target;

[0137]

[0138] In the formula, represents the information function, which is used to calculate the information between the j-th measurement and the t-th target. S jt (k) represents the information covariance corresponding to the information function. In the actual tracking situation, the value of B is related to the clutter density of the environment. If the clutter density is not very large, usually the constant B can be taken as 0.

[0139] In step 4, the combined information is calculated as follows:

[0140]

[0141] According to the Kalman filter, calculate the final state estimate value of the target t at time k and update the posterior covariance.

[0142] Z k|k = X k|k-1 + K k V t

[0143] P k∣k = P k∣k-1 - (1 - β 0t )K k S k K k T + K k (Z k|k - V t V t T )K k T

[0144] If step 1 is a failed track, then use X k|k-1 as the final state estimate value and input it into the next loop, and use the covariance matrix of the previous moment as the covariance matrix P k∣k .

[0145] Test data and algorithm verification:

[0146] In the embodiments of the present invention, a 77G millimeter-wave radar is used, and some parameters are as follows:

[0147]

[0148] As shown by Figure 7 , the scenario is a road scenario, and two pedestrians are moving crosswise. Figure 8 is the 2D point cloud data after radar signal processing. Through the radar multi-target tracking algorithm, the result graph of Figure 9 can be obtained. From the result graph, it can be seen that before and after the two pedestrians cross, the algorithm does not track the wrong target, and the track does not start repeatedly. The overall effect is good. There is no loss of accuracy due to simplifying the calculation process of JPDA, and the two pedestrians are not lost throughout the process, which verifies the effectiveness of the entire tracking method. As shown by Figure 10 , the scenario is a road scenario, with two bicycle targets and one pedestrian target. Among them, the two bicycles are moving in parallel, and the closest distance between them is less than 0.3 meters. According to Figure 12 the results shown, the clustering method in the preprocessing stage can effectively distinguish objects that are relatively close together, and there will be no track intersection in the subsequent data association part.

[0149] The above are only the preferred embodiments of the present invention and do not impose any formal restrictions on the present invention. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or make equivalent replacements for some of the technical features. Any modifications, equivalent replacements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A millimeter wave radar tracking method in a multi-target complex environment, characterized in that: The following steps are included: Step 1: Process the radio frequency signal received by the radar to generate millimeter wave radar point cloud data; Step 2: Define the radar range, remove the point cloud data beyond the range, and then cluster the point cloud using the DBSCAN clustering method to obtain the merge point; Step 3: Start the track and match the measured data with the track through data association gate judgment. If the association is successful, the target state is updated and the track is output. Otherwise, the scoreboard mechanism is used to execute track loss or maintenance operations.

2. The tracking method according to claim 1, characterized in that: The step 1 specifically includes the following sub-steps: Step 1.1: Separate the data of different channels in the RF signal, and improve the signal-to-noise ratio (SNR) of the signal by removing the DC component, filtering and noise reduction, and matching filtering; Step 1.2: Perform a fast Fourier transform of the signal in the distance dimension, and process the echo signal in the time domain using pulse compression technology to calculate the distance information of the target; Step 1.3: Perform Doppler FFT processing on the slow-time data to extract the Doppler frequency of the target, thereby obtaining the radial velocity of the target; Step 1.4: Remove static targets by using the moving target display MTI filtering technology; Step 1.5: Perform two-dimensional constant false alarm detection (CFAR) to screen out target points and their intensity information; Step 1.6: Estimate the direction of arrival (DOA) and obtain the target angle information.

3. The tracking method according to claim 1, characterized in that: In step 2, the radar range is defined and the point cloud data beyond the range is eliminated. Specifically, the starting origin, angle measurement range and distance measurement range are set, and the Euclidean distance between the measuring point and the starting origin and the angle are calculated to determine whether they are within the range to obtain the processed point cloud.

4. The tracking method according to claim 1, characterized in that: In step 3, the track is started using the m / n logic method, specifically: all measurements at the initial time are obtained, and the circular gates of the maximum distance and the minimum distance are established according to the estimated maximum speed and minimum speed and the sampling interval; Correlating the second frame measurement that falls within the initial correlation gate with the initial measurement to form a candidate track; The candidate track formed at the previous moment is extrapolated according to the Kalman filter KF algorithm, with the extrapolation point as the center, and the size of the subsequent gate is confirmed by the track extrapolation error covariance, so as to form the sampling confirmation area at the next moment. If the third frame measurement falls into the relevant gate, it is associated. If multiple measurements fall into it, the measurement closest to the extrapolation point is taken; Continue to extrapolate. If it is found that no measurement falls into the wave gate of the extrapolation point, the number of failures is recorded, and the extrapolation point is used as the predicted value of the current frame for extrapolation. If the number of failures is greater than nm times, it is proved to be a failure. Continue to extrapolate until a stable track is formed and the track is successfully started.

5. The tracking method according to claim 4, characterized in that: The Kalman filter KF algorithm specifically includes: The target state vector of the entire multi-target tracking process is expressed as: X k =[x,y,x',y'] Where x, y represent the position information of the target, and x', y' represent the speed information of the target; Prediction step: predict the state at the next moment based on the final prediction value at the previous moment; State transfer equation: X k|k-1 =FX k-1|k-1 +w k Prediction covariance matrix: P k|k-1 =FP k-1|k-1 F T +Q Where, X k-1|k-1 Indicates the final prediction value at the previous moment, w k is the process noise, Q is the process noise covariance matrix, and F is the state transfer matrix; Update step: adjust the predicted value of the prediction step according to the measured value to obtain the final state prediction value; Measurement equation: Z k =HX k +u k Measurement covariance matrix: S k =HP k|k-1 H T +R Kalman gain: Status Update: X k|k =X k|k-1 +K k (Z k -HX k|k-1 ) Update the covariance matrix: P k|k =(IK k H)P k|k-1 Where R is the measurement noise covariance matrix, Z k -HX k|k-1 For the information at this moment, the measurement matrix H is: The current updated state X k|k As the input for the next moment, the prediction and update process is repeated to achieve radar target tracking.

6. The tracking method according to claim 5, characterized in that: In step three, the measurement data is matched with the track through data association gate judgment, which is implemented by using the GJPDA algorithm. The algorithm specifically includes the following steps: Step 1: Use the Mahalanobis distance as a judgment, make a prediction based on the estimated value at the previous moment, and establish a wave gate with the center of the predicted value to obtain the effective measurement falling into the wave gate. Step 2: Obtain the confirmation matrix Ω from the relationship between the effective measurement and the target confirmation gate; Step 3: Calculate the interconnection probability β between the target and the measurement according to the likelihood function jt ; Step 4: Find the final combination information v k , according to the Kalman filter update step, the target state estimate and covariance matrix at the current moment can be obtained and wait for the next iteration.

7. The tracking method according to claim 6, characterized in that: In step 2, the establishment process of the confirmation matrix Ω is as follows: where w jt Indicates measurement j (j = 1, 2, ..., m k ) is associated with the target t (t = 0, 1, 2, ..., T), and its value is 0 or 1; when w jt = 0, it means that the jth measurement has no association with the tth target; when w jt =1, it means that the jth measurement is associated with the tth target; the first column of the confirmation matrix is ​​all 1, which means that at a certain moment, no target is associated with the measurement, and all measurements are false measurements; k is the number of measurements falling into the tracking gate at time k, and T is the number of targets.

8. The tracking method according to claim 6, characterized in that: In step 3, the likelihood function is used to replace the method of establishing multiple interconnection matrices and calculating the conditional probability of each interconnection matrix to obtain the interconnection probability of measurement j and target t. In this algorithm, the calculation formula of the final association probability is as follows: Where G jt It can be regarded as the probability that the jth measurement is associated with the tth target; In the formula represents the information function, which is used to calculate the information between the jth measurement and the tth target; S jt (k) represents the information covariance corresponding to the information function; in actual tracking situations, the value of B is related to the clutter density of the environment.

9. The tracking method according to claim 6, characterized in that: In step 4, the combined information is calculated as follows: According to the Kalman filter, the final state estimate of the target t at time k is calculated and the posterior covariance is updated. From k|k =X k|k-1 +K k In t P k∣k =P k∣k-1 -(1-β 0t )K k S k K k T +K k (WITH k|k -V t V t T )K k T If step 1 is a failed track, then X k|k-1 As the final state estimate, it is input into the next cycle, and the covariance matrix of the previous moment is used as the covariance matrix P of this moment. k∣k .

10. The tracking method according to claim 1, characterized in that: The scoreboard mechanism is specifically as follows: recording the number of times the target is lost and deciding whether to continue to maintain the track; if the number of times the target is lost exceeds a set threshold n, the target is considered to have been lost, a track loss operation is performed, and tracking of the track is terminated; otherwise, the track maintenance phase is entered, and subsequent measurement data is continued to be waited for to determine whether the target still exists.

Images