Multi-model filtering method and system based on target motion state recognition
By identifying the target's motion state and selecting the optimal filtering strategy, and employing methods such as Savitzky-Golay smoothing and weighted difference calculation, the problems of high computational complexity and insufficient response capability in maneuvering target tracking are solved, achieving efficient and robust target tracking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2025-12-18
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies for tracking maneuvering targets suffer from high computational complexity, insufficient tracking accuracy, and inadequate response capabilities, especially under conditions of high maneuverability where they struggle to quickly adapt to dynamic changes in the target.
By identifying the target's motion state and dynamically selecting the optimal filtering strategy, Savitzky-Golay smoothing filtering, weighted cumulative difference calculation, probabilistic decision-making, and combinations of different filtering models, including high-gain Alpha-Beta filtering, efficient and robust maneuvering target tracking is achieved.
It improves the ability to identify and respond to target maneuver patterns, reduces computational complexity, and enhances tracking accuracy and stability in complex environments.
Smart Images
Figure CN121880694A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar data processing technology, specifically to a multi-model filtering method and system based on target motion state recognition. Background Technology
[0002] Maneuvering target tracking is a core technology in modern radar, sonar, and navigation systems, and its performance directly affects the ability to continuously monitor targets, predict their trajectories, and analyze their behavior. In real-world applications, targets often do not maintain uniform linear motion but instead exhibit maneuvering behaviors such as acceleration, deceleration, and turning. These maneuvers cause a mismatch between the target motion model and the preset filtering model, leading to problems such as decreased tracking accuracy, filter divergence, and even track loss.
[0003] To address the problem of tracking maneuvering targets, traditional methods are mainly divided into two categories: single-model adaptive filtering algorithms and multi-model interactive algorithms. Single-model adaptive filtering algorithms, such as adaptive Kalman filtering and variable-dimensional filtering, adapt to target maneuvering by adjusting the noise covariance or model structure in real time. Although these methods have relatively low computational cost, they suffer from lag in model adjustment under strong maneuvering conditions, making it difficult to respond quickly to dynamic changes in the target. They are also susceptible to noise interference, resulting in insufficient tracking stability.
[0004] Multi-model interaction algorithms, especially interactive multi-model (IMM) algorithms, can cope with target mode switching to a certain extent by running multiple filters with different dynamic characteristics in parallel (such as uniform velocity CV models, uniform acceleration CA models, and coordinated turning CT models) and fusing states based on model probabilities. However, IMM algorithms have obvious limitations: first, the computational complexity increases significantly with the number of models, placing a heavy burden on real-time processing systems; second, the computational complexity of IMM algorithms is high when the number of models is large, and in some stable motion scenarios, multi-model interaction may introduce unnecessary computational overhead, which may affect filtering efficiency.
[0005] Therefore, how to reduce computational complexity while ensuring tracking accuracy, and enhance the ability to quickly identify and adaptively respond to target maneuvering patterns, has become a pressing technical challenge in this field. It is necessary to propose a novel maneuvering target tracking method that can achieve intelligent identification of motion states and dynamic optimization of filtering strategies, thereby enabling efficient and robust maneuvering target tracking in complex environments. Summary of the Invention
[0006] This invention proposes a method and system capable of identifying the motion state of a target and dynamically selecting the optimal filtering strategy.
[0007] The technical solution to achieve the purpose of this invention is as follows: This invention provides a multi-model filtering method based on target motion state recognition. The steps are as follows: First, the current track and associated point tracks are subjected to Savitzky-Golay smoothing filtering. The velocity and heading angle of each point on the smoothed track are calculated. Then, the weighted cumulative heading angle difference and velocity difference of the latest M points are calculated. Based on the weighted cumulative heading angle difference and velocity difference and the probability decision threshold configured by the system, the probabilities of the three models are calculated. The model with the highest probability is taken as the result of target motion state recognition, and the recognition result is updated to the current track state. Second, it is determined whether the associated point tracks fall within the wavelet gate. If they do not fall within the wavelet gate, high-gain Alpha-Beta filtering is used to filter the associated point tracks. If they fall within the wavelet gate, the corresponding filtering model is selected based on the current track state (target motion recognition result), and it is determined whether the state is the same as the previous moment. If they are the same, track filtering of the corresponding filtering model is directly performed. If they are different, the state vector X and covariance matrix P need to be initialized first. For IMM filtering, the model probability u also needs to be initialized. k Then, track filtering is performed; finally, the spatial instruction coordinate system information of the track is converted into radar array coordinate system information through the filtered data, and the filtered track is added to the aircraft list as a track point.
[0008] In addition, the present invention also provides a system for implementing the above method.
[0009] Compared with the prior art, the significant advantages of this invention are as follows: This invention is a filtering method oriented towards the target motion state. It selects the corresponding filter by identifying the target motion state. Compared with the filtering method of a single filtering model, it is more adaptable to the trajectory filtering of maneuvering targets. Compared with the interactive multi-model algorithm, it can show better comprehensive filtering effect when the target motion state is stable.
[0010] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0011] Figure 1 This is a flowchart illustrating the implementation of the present invention.
[0012] Figure 2 This is a flowchart for the target state recognition module. Detailed Implementation
[0013] The basic solution of this invention includes the following steps, and the specific processing flow is as follows: Figure 1 As shown.
[0014] like Figure 1 As shown, the trajectory filtering method provided by this invention mainly includes the following steps:
[0015] S1: To reduce the interference of measurement noise on state judgment, the x, y, and z coordinates of the input track and associated point tracks in the spatial rectangular coordinate system are first subjected to Savitzky-Golay filtering with a window length of 11 and an order of 2 to obtain a smoothed track, which is then output to S2. The Savitzky-Golay filter is used for this operation. This filter is a smoother based on local polynomial least squares fitting, which can effectively preserve the high-frequency variation characteristics of the signal and is superior to ordinary averaging filtering.
[0016] S2: Based on the results of S1, i.e., the smoothed track, the velocity and heading angle of each point in the smoothed track are calculated using the first-order difference method. The velocity and heading angle of the latest 9 points are taken, multiplied by individual weighting coefficients, and summed to obtain the weighted cumulative heading angle difference cMdd and the weighted cumulative acceleration difference cMdv, which are then output to S3. This invention proposes to calculate two weighted cumulative differences: the weighted cumulative heading angle difference cMdd and the weighted cumulative velocity difference cMdv at point M, and the calculation method is as follows.
[0017]
[0018] in, v represents the heading angle of the trajectory at time n. n This represents the velocity of the trajectory at time n. The larger cMdd and cMdv are, the closer the target is to turning motion and accelerating linear motion, respectively.
[0019] S3: The input weighted cumulative heading angle difference cMdd and weighted cumulative acceleration difference cMdv, along with their corresponding heading angle decision threshold ddth and velocity decision threshold dvth, are used to calculate the probabilities of three motion models (CV, CA, and CT) according to the probability calculation formula. The motion model with the highest probability is selected as the result of the current target motion state recognition and output to S5. The probability calculation formula is as follows:
[0020]
[0021] Where c i Let represent the probability of the i-th motion state.
[0022]
[0023] Among them, h Cv h CA h cT These represent the probability levels of motion states CV, CA, and CT, respectively; dvth and ddth represent the decision thresholds for cumulative velocity difference and cumulative heading angle difference, respectively; γCA and YCT represent the threshold offset coefficients for CA and CT state recognition, typically set to 1, but adjustable. γcT=1.2 to enable target motion state recognition to identify turning motions with smaller turning rates; the threshold offset coefficient is usually between 0.7 and 1.3.
[0024] S4: Based on the input gate information, the gate position and gate size can be obtained. Based on the associated point information, it is determined whether it falls within the small gate. If it does, proceed to S5; otherwise, it falls within the large gate, and proceed to S7. This invention uses two gates, large and small. If a point falls within the small gate, it is considered that its motion state has not changed significantly or the noise is low; if a point falls within the large gate, it is considered that its motion state has changed abruptly or that the prediction has deviated due to filter drift or that the noise is high.
[0025] S5: The associated point trace falls into the wavelet gate. Based on the input result of S3, i.e., the current trajectory state (target motion state recognition result), select the filtering model and perform filtering initialization as needed. This step includes:
[0026] S5.1: Based on the current track status (target motion status recognition result) from the S3 result, select the corresponding filtering model. For CV status, select CV-Kalman filtering; for CA status, select CA-Kalman filtering; for CT status, select IMM filtering which includes two CT-Kalman filters, and output to S5.2.
[0027] S5.2: Determine if the current track state is the same as the track state at the previous moment. If they are different, initialize the filtering model parameters of the track. The parameters to be initialized include: state vector X, error covariance matrix P, and IMM model probability u. k If they are the same, no initialization operation is performed. The selected filtering model is output to S6.
[0028] S6: Based on the filtering model input in S5, filter the input associated points. If the filtering model is CVKF or CAKF, execute the standard Kalman filtering procedure; if the filtering model is IMM, execute the Kalman-based IMM filtering procedure. Output the filtering result (track information) to S8. The filtering result includes: state vector X, error covariance matrix P, and IMM model probability u. k .
[0029] S7: The associated point trace falls into the large gate, undergoes high-gain Alpha-Beta filtering, and is used for uniform straight-line prediction of the input trajectory. The difference (information) between the predicted point and the associated point trace is calculated. The measurement points are filtered to obtain the filtered result X, where, Let K be the prediction point, and v be the filter gain vector and the innovation, respectively. K = [0.8 + 0.6 / 2.4], i.e., α = 0.8, β = 0.6, and the radar scan period T = 2.4 s. The filtering result (track information) is output to S8, and the filtering result includes the state vector X. A high-gain Alpha-Beta filter is applied to the current track. The high-gain Alpha-Beta filter actively reduces the reliability of the filtering model as the innovation gradually increases, primarily updating the track with measurement information. Before the measurement point gradually deviates from the prediction center and can no longer fall within the relevant gate, filtering performance is sacrificed to ensure the continuity and integrity of the track.
[0030] S8: Based on the inputs of S6 and S7, the x, y, and z coordinates of the target point in the spatial rectangular coordinate system are converted into the range, azimuth, and elevation angles in the radar coordinate system through spatial geometric relationships. The converted filter points are then added to the track, and finally the filtered track is output.
Claims
1. A multi-model trajectory filtering method and system based on motion state recognition, characterized in that, Includes the following steps: S1: Perform Savitzky-Golay smoothing filtering on the input track and associated point track, and output the smoothed track to S2; S2: Calculate the motion feature values of the smoothed input track using the first-order difference method based on its position information, and output the motion feature values to S3; S3: Calculate the probabilities of the three states based on the input motion feature values and the probability decision threshold configured by the system, take the state with the highest probability as the current motion state of the target, and output the current motion state of the trajectory to S5. S4: Determine the gate type into which the associated point falls based on the input gate information and associated point information, and output the gate type into which the associated point falls; if the associated point falls into a small gate, execute S5; if it falls into a large gate, execute S7. S5: The point falls into the wavelet gate. Select the corresponding filtering model for the current motion state of the input track and initialize the filtering according to the filtering needs. Output the initialized filtering model to S6. If no initialization is required, output the filtering model directly to S6. S6: Filter the input track and associated point information using the input filtering model to obtain the filtered track, and output the filtering result (track information) to S8. S7: The point trace falls into the large gate, and Alpha-Beta filtering is performed on the input track and associated point trace information to obtain the filtered track. The filtered result (track information) is output to S8. S8: Update the track points based on the input filtered track information, and output the filtered track.
2. The method according to claim 1, characterized in that, The smoothing filter described in step S1 uses Savitzky-Golay filtering with a window size of N and an order of order. The specific steps are as follows: the x, y, and z coordinates of the input track and the associated point track in the spatial rectangular coordinate system are respectively subjected to Savitzky-Golay filtering to obtain the smoothed track, which is then output to S2.
3. The method according to claim 1, characterized in that, The motion characteristic values mentioned in step S2 include: the weighted cumulative heading angle difference cMdd and the weighted cumulative acceleration difference cMdv. The weighted cumulative heading angle difference is the weighted cumulative value representing the heading angle differences of M consecutive points on the smoothed track, and the weighted cumulative acceleration difference represents the weighted cumulative value representing the velocity differences of M consecutive points on the smoothed track. Specifically, the steps are as follows: calculate the velocity and heading angle of each point on the smoothed track using the first-order difference method; take the velocity and heading angle of the latest M points; multiply them by individual weighting coefficients; and sum them to obtain the weighted cumulative heading angle difference cMdd and the weighted cumulative acceleration difference cMdv, which are then output to S3.
4. The method according to claim 1, characterized in that, Step S3, identifying the current motion state of the target, includes: calculating the probabilities of three motion models, including uniform linear motion (CV), uniformly accelerated linear motion (CA), and cooperative turning (CT); and selecting the state corresponding to the motion model with the highest probability as the current motion state. Specifically, the input weighted cumulative heading angle difference cMdd and weighted cumulative acceleration difference cMdv, along with their corresponding heading angle decision threshold ddth and velocity decision threshold dvth, are used to calculate the probabilities of the three motion models (CV, CA, and CT) according to the probability calculation formula. The motion model with the highest probability is selected as the result of the current target motion state identification and output to S5. The probability calculation formula is as follows: Where c i Let represent the probability of the i-th motion state. Among them, h Cv h CA h CT These represent the probability levels of motion states CV, CA, and CT, respectively; YCA and YCT represent the threshold offset coefficients for CA and CT state recognition.
5. The method according to claim 1, characterized in that, Step S4 involves determining the relevant gates into which the associated point trace falls. There are two types of gates: small gates and large gates. The size of the gates limits the distance, pitch angle difference, and velocity difference between the predicted and measured points in the XOY plane. Specifically, the steps are as follows: Based on the input gate information, the gate position and gate size can be obtained. Based on the associated point trace information, it is determined whether it falls within the small gate. If it does, proceed to S5; otherwise, it falls within the large gate, and proceed to S7.
6. The method according to claim 1, characterized in that, Step S5, after selecting the filtering model based on the current motion state, further includes: determining whether filtering initialization is needed; if so, initializing the filter parameters. The specific steps of S5 are: determining the current trajectory state (target motion state recognition result); if it is CV, selecting the CVKF filtering model; if it is CA, selecting the CAKF filtering model; if it is CT, selecting the IMM filtering model; secondly, determining whether the current trajectory state is the same as the previous state; if they are different, initializing the trajectory filtering model parameters. The parameters to be initialized include: state vector X, error covariance matrix P, and IMM model probability u. k If they are the same, no initialization operation is performed. The selected filtering model is output to S6.
7. The method according to claim 1, characterized in that, The filtering model mentioned in step S6 is the same as the filtering model selected in step S5. Specifically, the steps are as follows: Based on the filtering model input in S5, the input associated point traces are filtered. If the filtering model is CVKF or CAKF, a standard Kalman filtering procedure is executed; if the filtering model is IMM, a Kalman-based IMM filtering procedure is executed. The filtering result (track information) is then output to S8.
8. The method according to claim 1, characterized in that, The Alpha-Beta filtering described in step S7 is a high-gain Alpha-Beta filtering, where the filtering gain α > po and β > ve. The specific steps are: for the input trajectory, perform uniform straight-line prediction, calculate the difference (information) between the predicted point and the associated point trajectory, and based on... The measurement points are filtered to obtain the filtered result X, where, Let K be the prediction point, and K and v be the filter gain vector and innovation, respectively. Output the filtering result (track information) to S8.
9. The method according to claim 1, characterized in that, The specific steps described in step S8 are as follows: based on the filtering results input in step S6 or S7, and according to the spatial geometric relationship, the x, y, and z coordinates of the target point in the spatial rectangular coordinate system are converted into the range, azimuth, and elevation angles in the radar coordinate system, and the converted filtering points are added to the track. Finally, the filtered track is output.
10. A trajectory filtering system for implementing the method of any one of claims 1 to 9, comprising three modules, specifically: Target motion state recognition module: Receives external input tracks and associated points, recognizes the target motion state, and outputs the recognized motion state; Gate operation module: Receives gate information configured by the system and associated point traces input from the outside, determines the gate type into which the associated point trace falls, and outputs the determined gate type; if the associated point trace falls into a small gate, it also needs to receive the input target motion state, complete the selection and initialization of the filtering model, and output the filtering model. Track filtering and update module: Receives associated point track information and filtering model, performs track filtering, updates the track based on the filtering results, and outputs the filtered track.