A method and system for real-time tracking of radar flight targets
By combining deep learning scoring models and Markov chain Monte Carlo sampling, and utilizing the transformer target motion model, the problem of radar tracking speed and accuracy in strong clutter environments was solved, achieving fast and accurate target trajectory tracking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 四川启睿克科技有限公司
- Filing Date
- 2023-10-19
- Publication Date
- 2026-07-21
Smart Images
Figure CN117538858B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar tracking technology, specifically, to a method and system for real-time tracking of radar flying targets. Background Technology
[0002] The most fundamental and core task in multi-target tracking technology is to use the echo data from continuous wave radar to track the motion of multiple flying targets in real time and manage their motion tracks. The core of flight target tracking is to identify the real targets in the data and establish motion tracks for them. The process generally includes track initiation, track filtering, track maintenance, and track elimination. Track initiation specifically refers to the initial track establishment process before the target tracking system has determined that the target is moving stably. Track filtering refers to using certain rules or algorithms to judge and filter the reliability of established tracks, eliminating unreasonable and abnormal tracks. Track maintenance refers to matching the point data received by the radar in each cycle with the established tracks during real-time target tracking, updating the track information to obtain the latest target track. The track maintenance process generally involves describing the target's motion state and selecting a data matching algorithm. Track elimination refers to deleting the corresponding tracks when the target flies out of the radar detection range, stops moving for a long time, or is lost to radar tracking for more than a certain period of time, thereby ensuring the accuracy of target tracking results and the real-time nature of tracking speed.
[0003] Traditional track initiation methods can be categorized into sequential processing methods and batch processing methods based on different data processing approaches. Sequential processing methods include logical methods and heuristic rule-based methods; batch processing methods include Hough transform and its variants. The advantage of sequential processing methods is low computational cost and fast processing speed, but the disadvantage is that the accuracy in establishing correct tracks is generally limited. Since sequential processing methods cannot distinguish clutter, increased clutter will directly interfere with track formation; therefore, this method is typically suitable for environments with weak clutter. Batch processing methods, on the other hand, involve high computational cost and long processing time, but offer good results and can be used in environments with strong clutter.
[0004] Traditional track selection methods include logical and rule-based methods, which are generally the same as the logical and rule-based methods for track initiation. Therefore, in some classic target tracking frameworks, the track initiation process and the track selection process are generally combined into the track initiation process.
[0005] Track maintenance involves track matching and target kinematics simulation. Traditional methods for target kinematics simulation include the use of a family of Kalman filters (classical Kalman filter, extended Kalman filter, unscented Kalman filter, etc.). More recent methods include Interacting Multiple Model (IMM) and Bernoulli family filters (e.g., PMBM (Poisson-Bernoulli hybrid filter), δ-GLMB (δ-generalized label-multiple Bernoulli filter), etc.). Traditional Kalman filtering is advantageous due to its fast computation speed; while IMM and Bernoulli family filters can adapt to more complex motion scenarios. However, both of these methods require setting a large number of parameters in advance, making them inconvenient to use. Classic track matching algorithms (also called data association algorithms) include Nearest Neighbor (NN), Global Nearest Neighbor (GNN), Probabilistic Data Association (PDA), Joint Probabilistic Data Association (JPDA), Multiple Hypothesis Tracking (MHT), and dynamic programming. Nearest neighbor tracking algorithms are fast but generally ineffective; while probabilistic correlation tracking algorithms and multi-hypothesis tracking algorithms are more effective but more complex, time-consuming, or impossible to solve.
[0006] The environment for tracking flying targets with modern radar is becoming increasingly challenging. On one hand, strong clutter is becoming more prevalent in the objective environment, and the number of various targets in space (such as birds, multi-rotor drones, fixed-wing drones, and small aircraft) is also increasing, all of which increase the difficulty of tracking. On the other hand, current application scenarios demand increasingly higher speed and accuracy in target tracking, making traditional flight target tracking solutions increasingly inadequate. While some methods in the field of target tracking have initially utilized deep learning models for certain tasks, these methods still face many problems in practical applications, which can be summarized as follows:
[0007] 1. Currently, radar detection faces complex and cluttered environments. When faced with a large amount of echo data, traditional flight target tracking methods (such as JPDA, MHT, etc.) suffer from excessive computation, resulting in a decrease in tracking speed.
[0008] 2. Currently, rule-based track initiation methods are generally ineffective, while deep learning-based track initiation methods use manually designed features, which have limitations. At the same time, deep learning models do not utilize the temporal characteristics of target motion data, resulting in an inability to learn the target's motion characteristics well. They are not suitable for targets with complex motion characteristics and have low tracking accuracy. Summary of the Invention
[0009] The purpose of this invention is to provide a method and system for real-time tracking of radar flying targets, which solves the problems of reduced tracking speed in strong clutter environments when radar performs multi-flying target tracking tasks, and the limitations of deep learning-based track initiation methods which use manually designed features; in addition, the lack of utilization of the temporal characteristics of target motion data leads to low tracking accuracy.
[0010] The present invention solves the above problems through the following technical solution:
[0011] A method for real-time tracking of radar flying targets includes:
[0012] Step A: Receive the echo point data of the current radar frame and perform preprocessing;
[0013] Step B: Predict the motion state of each established target track, use the prediction results and preprocessed data to perform track matching, and establish new target tracks for unmatched data;
[0014] Step C: Use a deep learning scoring model to score all current tracks and remove tracks with scores below the threshold.
[0015] Step D: Iteratively calculate the optimal hypothetical trajectory in the multi-hypothesis framework using a Markov chain Monte Carlo sampling method;
[0016] Step E: Manage the existing hypothetical tracks based on the obtained optimal hypothetical tracks;
[0017] Step F: Post-process the target tracking results and output the current target tracking results.
[0018] Furthermore, the preprocessing in step A includes removing radar clutter and aggregating the split radar clutter data.
[0019] Furthermore, the prediction of the motion state of each established target trajectory specifically includes:
[0020] The sequence data consisting of the coordinates of the track segment formed by the N time points preceding the current time point of the target track;
[0021] Feature vectors are obtained by extracting features from sequence data;
[0022] The feature vector is encoded to obtain the encoded vector. The encoded vector and the feature vector are then spatially encoded to obtain the spatial information vector. Finally, the encoded vector and the spatial information vector are decoded together to obtain the prediction result, which is the position coordinates of the target track at the current time.
[0023] Furthermore, the prediction of the motion state of each established target trajectory is accomplished by a feature extraction module and a target motion model, wherein:
[0024] The feature extraction module is used to extract feature vectors from sequence data;
[0025] The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. The feature encoding module encodes the feature vectors to obtain the encoded vector. The spatial perception module encodes the encoded vector and the feature vectors to obtain the spatial information vector. The decoding module decodes the spatial information vector and the encoded vector together to obtain the prediction result.
[0026] Furthermore, the feature encoding module and the decoding module are each composed of multiple stacked transformers, and the spatial perception module is composed of a fully connected layer, a 3D sparse convolutional layer and a softmax layer stacked together.
[0027] Furthermore, the process of matching the predicted results and preprocessed data for track matching involves: matching the predicted results with the preprocessed data at the current time point using a gate method; at each time point, assigning the successfully matched point data to the established target track; and the process of associating the target track with the echo data to update the target track, thereby updating its state. The gate shape is rectangular or elliptical, and the gate distance uses Euclidean distance or Mahalanobis distance.
[0028] Furthermore, the deep learning scoring model is composed of fully connected layers, convolutional layers, and softmax layers stacked together. The input data is a sequence of feature data of each target trajectory from the previous N time points starting from the current time point. The feature data of the target trajectory at each time point consists of the target's three-dimensional coordinates, signal-to-noise ratio, echo amplitude, velocity, and acceleration at the current time point.
[0029] Furthermore, the method based on Markov chain Monte Carlo sampling specifically includes the construction of Markov chains and the construction of the target hypothetical trajectory sampling distribution. The optimal score of each hypothetical trajectory is obtained through Markov chain Monte Carlo sampling, and then hypothetical trajectories with optimal scores greater than a threshold are selected.
[0030] Furthermore, the management of existing hypothetical tracks in step E includes: updating the status information of hypothetical tracks, deleting hypothetical tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks in the hypothesis tree that exceed a threshold number.
[0031] Furthermore, the target tracking result in step F is a trajectory sequence consisting of all coordinates from the target's starting point to the current time.
[0032] A system for real-time radar target tracking includes:
[0033] The point data preprocessing module is used to preprocess the echo data of the continuous wave radar. The preprocessing includes clutter removal and aggregation of split point data.
[0034] The hypothesis tree module is used to construct a hypothesis tree from all tracks starting from the same origin. The hypothesis tree records the coordinates, track number, score, length, number of missing points, and maximum consecutive missing value of the track.
[0035] The feature extraction module is used to extract the kinematic feature vector of the target trajectory by taking the sequence data of the three-dimensional coordinates of the target trajectory from the previous time point to N time points in advance.
[0036] The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. It takes the extracted kinematic feature vector as input and outputs the predicted position coordinates of the target trajectory at the current moment.
[0037] The track matching module uses the gate method to match the target track with all echo data received at the current time point in the preprocessed data. New tracks are established for the target track and each echo data falling within the gate. The process of associating the target track with the echo data is essentially updating the target track. If multiple new tracks are formed, they are all updated at the next time step.
[0038] The New Track Initiation Module is used to build new track trees and tracks from the preprocessed data that failed to match, according to the rules of the MHT tracking framework.
[0039] The track scoring feature extraction module is used to extract track scoring features from the historical data of the target track. The historical data includes coordinates, echo amplitude, signal-to-noise ratio, and computational cost of constructing the track from the historical data.
[0040] A deep learning scoring model is used to take track scoring features as input and output the track's score value.
[0041] The Markov Chain Monte Carlo Sampling Module is used to iteratively calculate the best score of all current hypothetical target tracks using the MCMC sampling algorithm or the Gibbs sampling algorithm.
[0042] The track management module is used to manage existing tracks, including updating track status information, deleting target tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks in the hypothesis tree that exceed a threshold number.
[0043] The target tracking result output module extracts the historical coordinate sequence and corresponding track number from the target track information and then outputs them to the display terminal.
[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0045] (1) In the face of environments with strong clutter, the present invention can find the target from the received radar echo point data, track the target in real time, and provide real-time track information of the tracked target; it has a faster tracking speed when the amount of data is large; by evaluating the rationality of the target track, the accuracy of the target track is improved, and the number of assumed tracks is reduced, which also improves the tracking speed.
[0046] (2) The present invention is adapted to strong clutter environment and can capture and continuously track flying targets in such environment.
[0047] (3) The present invention uses a single transformer-based target motion model to predict the trajectory of all targets, eliminating the need to build a separate model for each target, thus reducing the number of parameters and improving accuracy. Attached Figure Description
[0048] Figure 1 This is a flowchart of the present invention;
[0049] Figure 2 This is a system block diagram of the present invention. Detailed Implementation
[0050] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.
[0051] Example:
[0052] Combined with appendix Figure 1 As shown, a method for real-time tracking of a radar flying target includes:
[0053] Step A: Receive the echo point data of the current radar frame and perform preprocessing; preprocessing includes radar clutter point removal based on distribution estimation and radar point aggregation based on clustering algorithms (such as DBSCAN clustering, hierarchical clustering, etc.).
[0054] Step B1: Predict the motion state of each established target trajectory using the same transformer-based target motion model, including:
[0055] 1) Sequence data consisting of the coordinates of the track segments formed by the N time points preceding the current time point of the target track;
[0056] 2) Extract features from sequence data to obtain feature vectors;
[0057] 3) Encode the feature vector to obtain the encoded vector, encode the encoded vector and the feature vector to obtain the spatial information vector, and then decode the encoded vector and the spatial information vector together to obtain the prediction result, that is, the position coordinates of the target track at the current time.
[0058] Preferably, the prediction of the target trajectory motion state is jointly completed by the feature extraction module and the target motion model. The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. The input data of the feature extraction module is the sequence data of all three-dimensional coordinates of the target trajectory from the current time period to the N time periods before it. N is selected as needed, generally 3 to 5. After the feature extraction module extracts the feature vector, it is then input into the feature encoding module of the target motion model to obtain the encoded vector. The encoded vector and the input data are then input into the spatial perception module for spatial information encoding to obtain the spatial information vector. Next, the encoded vector and the spatial information vector feature are input into the decoding layer for decoding, and finally into the output layer to obtain the result. Preferably, the feature encoding module is stacked 8 to 10 layers, with each layer outputting an encoded vector of length 56 or 128; the decoding module is also stacked 8 to 10 layers, with each layer outputting an encoded vector of length 56 or 128. The spatial perception module outputs M vectors of length 56 or 128, where M is pre-set to represent the number of possible targets, generally set to 100 or 150.
[0059] Step B2: Use the predicted results and preprocessed data to perform track matching, and establish new target tracks for the unmatched preprocessed data;
[0060] Track matching refers to assigning successfully matched point data obtained by the gate method to an established target track at each time point, thereby updating the state of the target track. Specifically, the method involves matching the motion state prediction results of the target track using a target motion model with the preprocessed data at the current time point. The gate method is chosen for matching; the gate shape can be rectangular or elliptical, and the gate distance can be Euclidean distance or Mahalanobis distance; the gate size needs to be set in advance.
[0061] Preprocessed data that fails to match existing tracks are used to create new tracks according to certain rules. This method generally adopts the MHT framework; therefore, new hypothesis trees and hypothesis tracks are built based on the MHT framework, and new tree numbers and track numbers are assigned.
[0062] Step C: Use a deep learning scoring model to score all current tracks and remove tracks with scores below the threshold.
[0063] Preferably, the deep learning scoring model consists of stacked fully connected layers, convolutional layers, and softmax layers. The deep learning scoring model outputs a score for each track, with the score ranging from [0,1]. The scoring threshold needs to be set in advance. The input data for the deep learning scoring model is constructed according to the following process:
[0064] First, extract the sequence data of the target trajectory from the current time point forward T time points; T is selected in advance as needed, and can be 3 or 5. Next, construct the input data, which consists of two concatenated parts. The first part is a vector composed of features extracted from the sequence data at each time point. The features at each time point specifically include: three-dimensional Euclidean coordinates, signal-to-noise ratio, echo amplitude, etc. The second part contains motion information such as velocity, acceleration, and yaw angle of the sequence.
[0065] The loss function of the target motion model consists of multiple parts. The overall loss function is a combination of the prediction loss of track coordinates and the matching loss of multi-target collective prediction. The prediction loss of track coordinates uses a smooth L1 loss function, and the matching loss uses a constructed matching loss function.
[0066] The loss function used in the deep learning scoring model is L2 loss.
[0067] During training, the corresponding model is optimized by maximizing the loss function of each model. During training, when the loss value is not within a reasonable range, the model parameters are adjusted and training continues until the loss value drops to a reasonable range. The corresponding model is then used as the final target motion model and deep learning scoring model.
[0068] The number of neural network layers and neurons per layer in each sub-module network of the model, the setting of the number of stacked sub-module layers, the setting of the feature vector length, the initialization of the parameters of each layer in the model, the connection and alignment between each network layer, and the selection of the model training optimizer and parameters can be set according to the actual situation.
[0069] Step D: Iteratively calculate the optimal hypothetical trajectory in the multi-hypothesis framework using a Markov chain Monte Carlo sampling method;
[0070] The Markov chain Monte Carlo sampling algorithm can choose Metropolis-Hastings sampling or Gibbs sampling. The number of iterations, Markov chain parameters, number of samplings, and threshold for the optimal trajectory all need to be set in advance.
[0071] Step E: Manage the existing hypothetical tracks based on the obtained optimal hypothetical tracks;
[0072] Track management specifically includes: updating track status information, deleting tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks from the hypothesis tree that exceed a threshold number. Track status information includes track number, track length, track coordinates, whether it is a spurious node, whether the track has ended tracking, the track score, and the number of missing points. When a track contains more than a threshold of consecutive missing values, the track will be deleted. The length of consecutive missing values needs to be set in advance, such as 3. There is a threshold limit to the number of hypothetical tracks that each hypothesis tree can contain, such as 7 to 10. Hypothetical tracks exceeding the threshold and with the lowest score will be deleted.
[0073] Step F: Post-process the target tracking results and output the current target tracking results. Post-processing refers to extracting the historical coordinate sequence and corresponding track number of the target tracking result from the hypothetical track.
[0074] This invention uses a unified target motion model, eliminating the need to set motion rules for different motion modes and providing more accurate target motion state prediction capabilities. It also eliminates the need to build a motion model for each trajectory, reducing the number of models and parameters in the target tracking process. Furthermore, the deep learning scoring model used evaluates the target's trajectory, eliminating false trajectories and reducing the complexity of subsequent calculations; and Markov chain Monte Carlo sampling can handle scenarios with large datasets. These points demonstrate the strong adaptability and wide applicability of this method.
[0075] Example 2:
[0076] Combined with appendix Figure 2 As shown, a system for real-time tracking of radar flying targets includes:
[0077] The point data preprocessing module is used to preprocess the echo data of the continuous wave radar: first, clutter is removed, and then the split point data is aggregated to obtain the standard input data for the subsequent tracking process.
[0078] The hypothesis tree module is used to construct a hypothesis tree from all tracks starting from the same point. The hypothesis tree records various information about the track, such as coordinates, track number, score, length, number of missing points, and maximum consecutive missing value.
[0079] The feature extraction module is used to extract the kinematic feature vector of the target trajectory by taking the sequence data of the three-dimensional coordinates of the target trajectory from the previous time point to N time points in advance.
[0080] The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. It takes the extracted kinematic feature vector as input and outputs the predicted position coordinates of the target trajectory at the current moment.
[0081] The track matching module uses the gate method to match the target track with all echo data received at the current time point in the preprocessed data, and establishes new tracks for the target track and all echo data falling within the gate.
[0082] The new track initiation module is used to build new track trees and tracks based on the rules of the multi-hypothesis MHT tracking framework from the preprocessed data that failed to match. The track scoring feature extraction module is used to extract track scoring features from the historical data of the target track. The historical data includes coordinates, echo amplitude, signal-to-noise ratio, and computational complexity constructed from the historical data.
[0083] The deep learning scoring model consists of fully connected layers, convolutional layers, and softmax layers. It takes track scoring features as input and outputs the track's score value.
[0084] The Markov Chain Monte Carlo Sampling Module is used to iteratively calculate the best score of all current hypothetical target tracks using the MCMC sampling algorithm or the Gibbs sampling algorithm.
[0085] The track management module is used to manage existing tracks, including updating track status information, deleting target tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks in the hypothesis tree that exceed a threshold number.
[0086] The target tracking result output module extracts the historical coordinate sequence and corresponding track number from the target track information and then outputs them to the display terminal.
[0087] A real-time radar target tracking system integrating deep learning scoring models and Markov chain Monte Carlo sampling can quickly detect and continuously track the flight path of a target in a cluttered environment, given the echo data of the current frame of a continuous wave radar, without relying on manually defined tracking rules or other complex information.
[0088] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.
Claims
1. A method for real-time tracking of a radar-flying target, characterized in that, include: Step A: Receive the echo point data of the current radar frame and perform preprocessing; Step B: Predict the motion state of each established target track, use the prediction results and preprocessed data to perform track matching, and establish new target tracks for unmatched data; Step C: Use a deep learning scoring model to score all current tracks and remove tracks with scores below the threshold. Step D: Iteratively calculate the optimal hypothetical trajectory using a Markov chain-based Monte Carlo sampling method; Step E: Manage the existing hypothetical tracks based on the obtained optimal hypothetical tracks; Step F: Post-process the target tracking results and output the current target tracking results; Step B uses the same transformer-based target motion model to predict the motion state of each established target track, specifically including: The sequence data consisting of the coordinates of the track segment formed by the N time points preceding the current time point of the target track; Feature vectors are obtained by extracting features from sequence data; The feature vector is encoded to obtain the encoded vector. The encoded vector and the feature vector are then spatially encoded to obtain the spatial information vector. Finally, the encoded vector and the spatial information vector are decoded together to obtain the prediction result, which is the position coordinates of the target track at the current time.
2. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The preprocessing in step A includes removing radar clutter and aggregating the split radar clutter data.
3. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The prediction of the motion state of each established target trajectory is accomplished by a feature extraction module and a target motion model, wherein: The feature extraction module is used to extract feature vectors from sequence data; The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. The feature encoding module encodes feature vectors to obtain encoded vectors. The spatial perception module encodes the encoded vectors and feature vectors to obtain spatial information vectors. The decoding module decodes the spatial information vectors and encoded vectors together to obtain the prediction result.
4. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The process of matching the predicted results and preprocessed data for track matching involves matching the predicted results with the preprocessed data at the current time point, assigning the matched track data to the established target track at each time point, thereby updating the status of the target track.
5. The method for real-time tracking of a radar flying target according to claim 4, characterized in that, The method for matching the prediction results with the preprocessed data at the current time point is the gate method. The gate shape is rectangular or elliptical, and the gate distance is Euclidean distance or Mahalanobis distance.
6. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The method based on Markov chain Monte Carlo sampling specifically includes the construction of Markov chains and the construction of the target hypothetical trajectory sampling distribution. The optimal score of each hypothetical trajectory is obtained through Markov chain Monte Carlo sampling, and then hypothetical trajectories with optimal scores greater than a threshold are selected.
7. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The management of existing hypothetical tracks in step E includes: updating the status information of hypothetical tracks, deleting hypothetical tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks in the hypothesis tree that exceed a threshold number.
8. The method for real-time tracking of a radar flying target according to claim 1, characterized in that, The target tracking result in step F is a trajectory sequence consisting of all coordinates from the target's starting point to the current time.
9. A system for real-time tracking of radar flying targets, characterized in that, include: The point data preprocessing module is used to preprocess the echo data of the continuous wave radar. The preprocessing includes clutter removal and aggregation of split point data. The hypothesis tree module is used to construct a hypothesis tree from all tracks starting from the same origin. The hypothesis tree records the coordinates, track number, score, length, number of missing points, and maximum consecutive missing value of the track. The feature extraction module is used to extract the feature vector of the target trajectory by taking the sequence data of the three-dimensional coordinates of the target trajectory starting from the time node before the current time node and going back N time nodes. The target motion model consists of a feature encoding module, a decoding module, and a spatial perception module. The feature encoding module encodes the input feature vector to obtain an encoded vector. The spatial perception module encodes the encoded vector and the feature vector to obtain a spatial information vector. The decoding module decodes the spatial information vector and the encoded vector together to obtain the prediction result and outputs the predicted position coordinates of the target trajectory at the current time. The track matching module uses the gate method to match the target track with all echo data received at the current time point in the preprocessed data, and establishes new tracks for the target track and all echo data falling within the gate. The new track initiation module is used to build new track trees and tracks from preprocessed data that failed to match; The track scoring feature extraction module is used to extract track scoring features from the historical data of the target track. The historical data includes coordinates, echo amplitude, signal-to-noise ratio, and computational cost of constructing the track from the historical data. A deep learning scoring model is used to take track scoring features as input and output the track's score value. The Markov chain Monte Carlo sampling module is used to iteratively calculate the best score of all current hypothetical target tracks through a sampling algorithm; The track management module is used to manage existing tracks, including updating track status information, deleting target tracks with consecutive missing values exceeding a threshold, and deleting hypothetical tracks in the hypothesis tree that exceed a threshold number. The target tracking result output module extracts the historical coordinate sequence and corresponding track number from the target track information and then outputs them to the display terminal.