An agimm tracking method based on maneuver detection ranking

Abstract

This invention discloses an AGIMM tracking method based on maneuver detection and ranking, comprising the following steps: S1, establishing a cooperative turning tracking model; S2, updating the turning rate of the cooperative turning tracking model using the AGIMM algorithm; S3, discriminating target maneuvers; S4, calculating the discrete turning rate using a discrete turning rate calculation model; S5, adjusting the model structure; and S6, adjusting the model probability transition matrix. This invention proposes a target tracking method for consumer-grade small unmanned aerial vehicles (UAVs) based on the AGIMM algorithm. By combining a target maneuver discrimination method with a confidence ranking strategy, it improves the convergence rate of the AGIMM algorithm for turning rates and its ability to handle disturbances, effectively enhancing the target tracking efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of aviation, and more particularly to an AGIMM tracking method based on maneuver detection and sorting. Background Technology

[0002] With the development of drone technology, the spectrum of drone systems is no longer limited to military applications; its applications cover people's production, daily life, consumption, and entertainment, and low-threshold drones are increasingly entering the public eye. However, the widespread use of drones has also brought significant risks to airspace safety, especially near major airports, where incidents of large-scale flight evacuations and diversions due to unauthorized drone use are frequent. Cases of successfully apprehending the perpetrators afterward are rare, mainly due to the lack of low-altitude monitoring and surveillance measures near important targets, making continuous tracking and surveillance of illegally flying targets impossible. Therefore, developing tracking technologies for "low, slow, and small" targets, especially tracking and verification technologies for small drones, is of great practical significance for improving the operational safety of low-altitude airspace.

[0003] Small unmanned aerial vehicles (UAVs) operate at slower speeds and are lightweight, but possess high maneuverability, exhibiting significantly different motion characteristics compared to conventional manned aircraft or missiles. Conventional aircraft, besides linear motion, exhibit pronounced circular turning characteristics during maneuvers. However, some small UAVs, due to their light weight and high maneuverability, have smaller turning radii and more flexible maneuvers. Current target tracking methods can be categorized into single-model algorithms and multi-model algorithms. Single-model algorithms primarily estimate the trajectory of moving targets based on models of uniform motion and uniform acceleration, Coordinated Turn (CT) models, Singer models, current statistical models, and Jerk models. However, single-model algorithms are only effective when the target strictly adheres to the assumed motion pattern; otherwise, target loss can occur, significantly limiting their application. In 1965, Magil first proposed the Multiple Model (MM) algorithm to address the shortcomings of single-model algorithms, but these models, lacking interaction with each other, still exhibited significant errors during tracking. In 1988, Blom Henk first proposed the Interactive Multiple Model (IMM) algorithm, which has been widely used in target tracking research. However, the IMM algorithm is still limited by the selected model set and lacks adaptability, especially for tracking turning targets. To improve the algorithm's adaptability, a series of improved algorithms based on the IMM algorithm have been proposed, achieving better target tracking results. LI XR et al. proposed a variable-structure IMM algorithm that can adaptively adjust the model set, reducing the computational cost of multiple models while improving tracking accuracy. QIAO XD proposed an Adaptive Grid IMM (AGIMM) algorithm to address the deficiency of the cooperative turning model's turning rate heavily relying on prior knowledge. This algorithm can continuously adjust the turning rate during tracking, improving the tracking accuracy of maneuvering targets. Zhu Hongfeng used a BP neural network to train the turning model in the IMM algorithm to discriminate the turning rate, thereby achieving adaptive tracking of the turning rate. Shao Kun improved the convergence speed of the turning rate of the AGIMM algorithm based on fuzzy logic, effectively improving the tracking efficiency of the original AGIMM algorithm. Pan Meimei proposed a variable structure AGIMM algorithm based on maneuver discrimination, which can adaptively change the model probability transition matrix and is used to solve the tracking problem of hypersonic targets in near space.

[0004] In summary, horizontal maneuvering tracking and monitoring of low-altitude, slow-moving, and small aerial targets remains a significant challenge for low-altitude airspace management. Traditional multi-model algorithms can effectively improve target tracking accuracy through model interaction, but their accuracy is limited for targets undergoing continuous maneuvers because the target's maneuverability cannot be predicted. The AGIMM algorithm addresses the adaptive tracking problem based on turn rate, but its tracking accuracy is affected by disturbances. The IMM algorithm is widely used in aerial target tracking, and model switching during maneuvers significantly impacts tracking performance. However, traditional IMM algorithms are not ideal for adaptive maneuvers, particularly for lightweight and highly maneuverable targets like UAVs, where significant tracking errors occur. Therefore, it is necessary to research a tracking method to improve tracking accuracy and address these issues. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing an AGIMM tracking method based on maneuver detection and ranking to improve the robustness of target tracking.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] An AGIMM tracking method based on motion detection and sorting includes the following steps:

[0008] S1. Establish a cooperative turning tracking model;

[0009] S2. Update the turning rate of the cooperative turning tracking model using the AGIMM algorithm;

[0010] S3. Identify target maneuvers;

[0011] S4. Calculate the discrete turning rate using the discrete turning rate calculation model;

[0012] S5. Adjust the model structure;

[0013] S6. Adjust the model probability transition matrix.

[0014] Furthermore, in step S1, assuming the UAV performs a turning maneuver in the horizontal plane, the velocity scalar remains constant during the turn, and the angular velocity is expressed as... Therefore, the calculation formula for the cooperative turning tracking model is:

[0015] In the formula: V represents the spatial position and velocity of the target; V(k) is the discrete-time white noise sequence of the system. The covariance matrix of the noise is: In the formula: when ω>0, the target turns left; when ω<0, it turns right; when ω=0, it becomes a uniform motion model; q is the system noise variance.

[0016] Furthermore, in step S2, the turning rates of the three cooperative turning tracking models are arranged from smallest to largest, specifically as follows: ,and ,Will As the turning rate of the left jump model. As an intermediate model turning rate As the turning rate of the right jump model; the posterior probability corresponding to each turning rate is: After each round of filtering is completed, the turning rates are updated according to the corresponding posterior probabilities.

[0017] Furthermore, in step S3, the determination of the target maneuver includes the following steps: S31. After the k-th filtering, the three cooperative turning tracking models updated by the AGIMM algorithm yield three residual vectors respectively. These correspond to three posterior probability values. Using the residual of the cooperative turning tracking model i with the largest posterior probability value as the criterion, its distance function is calculated using the following formula: In the formula: The residual of the cooperative turning tracking model i at time k; Let i be the residual covariance of the collaborative turning tracking model i at time k; S32. Set the threshold value M for maneuvering and turning; if If the condition is met, it is determined that a maneuver has occurred; otherwise, it is determined that no maneuver has occurred.

[0018] Furthermore, in step S4, the calculation formula for the discrete turning rate calculation model is as follows: ;

[0019] Furthermore, step S5, adjusting the model structure, includes the following steps: S51, when At this point, the turning rate calculated by the discrete turning rate calculation model is used as the main factor. The turning rates are sorted from largest to smallest according to the various posterior probability values, which are as follows: Its corresponding turning rate

[0020] In the formula: α is the adjustment factor; Minimum grid spacing; S52, when At this point, the turning rate of the left-jump model is the primary factor, and the models are sorted according to their posterior probability values; if The system is determined to have engaged in maneuvering; the model spacing adjustment parameters are as follows: S53, when At this point, the intermediate model's turning rate is the primary factor, and the models are sorted based on their posterior probability values; if If a maneuver occurs, the model spacing adjustment parameter is determined to be: S54, when At this point, the right-jump model's turning rate is the primary factor, and the models are sorted based on their posterior probability values; if The system is determined to have engaged in maneuvering; the model spacing adjustment parameters are as follows:

[0021] Furthermore, in step S6, the formula for calculating the model probability transition matrix is: In the formula: Let be the likelihood function of model j after filtering at time k; Let be the transition probability from model i to model j at time k;

[0022] After improvement:

[0023] Compared with the prior art, the advantages and positive effects of this invention are:

[0024] This invention proposes a target tracking method for consumer-grade small UAVs based on the AGIMM algorithm. It combines a target maneuver discrimination method and a confidence ranking strategy to improve the convergence rate of the AGIMM algorithm for turning rate and its ability to cope with disturbances, effectively improving the tracking efficiency and accuracy of the target. Furthermore, it incorporates a discrete turning rate calculation model and, through ranking, promptly eliminates models with large residuals, which can make the model more robust and further improve the effectiveness of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of discrete point extraction.

[0027] Figure 2 A flight path diagram for drone operations;

[0028] Figure 3 Aerial flight path map of the drone;

[0029] Figure 4 A comparison chart showing the tracking status of the angular velocity of the flight trajectory;

[0030] Figure 5 shows a comparison of the tracking performance of flight trajectory one; where a is a schematic diagram in the X direction; b is a schematic diagram in the Y direction; and c is a schematic diagram in the Z direction.

[0031] Figure 6 This is a comparison chart showing the tracking status of the two angular velocities of the flight trajectory. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.

[0033] This invention relates to an AGIMM tracking method based on motion detection and sorting; its specific operation steps are as follows:

[0034] 1. Cooperative turning tracking model

[0035] Tracking turning maneuvers has always been a challenge in target tracking. The AGIMM algorithm, based on a cooperative turning model, can adaptively adjust the estimation of the target's turning rate. For some small and medium-sized UAVs, their maneuverability doesn't resemble that of manned combat aircraft, which perform loops or spirals. Most of their maneuvers are horizontal turns, but the turning rate varies considerably. We can assume the UAV is performing a turn in a horizontal plane, with the velocity scalar remaining constant during the turn, and the angular velocity... The discretized state equation can then be expressed as: (1)

[0036] In the formula: V represents the spatial position and velocity of the target; V(k) is the discrete-time white noise sequence of the system.

[0037] The covariance matrix of the noise is:

[0038] In the formula: when ω>0, the target turns left; when ω<0, it turns right; when ω=0, it becomes a uniform motion model; q is the system noise variance.

[0039] The low-altitude detection platform can be a traditional Doppler radar, an infrared / electro-optical radar, or an ADS-B-based monitoring platform. However, the ADS-B system is a cooperative surveillance system and cannot effectively track and monitor some non-cooperative targets. Therefore, this invention uses a traditional Doppler radar as an example to construct a measurement model, assuming the measurement equation is:

[0040]

[0041] 2. AGIMM Algorithm

[0042] The AGIMM algorithm is built upon multiple turning models, each with a specific turning characteristic.

[0043] 3. MDS-AGIMM Algorithm

[0044] Traditional AGIMM algorithms rely on a certain amount of historical residual data for target turning rate tracking, resulting in slow convergence. The higher the turning rate threshold, the longer the search time. If the target maneuvers continuously with significant differences in turning rates, large position estimation errors and target loss can occur. Turn rate calculation methods based on maneuver discrimination determine whether the target has maneuvered by calculating the residual information of point traces. If a maneuver is detected, a subset of discrete point traces is extracted, and the turning rate of these discrete points is used to predict the target's current turning rate. However, this method also carries the risk of misjudging maneuvers. Therefore, combining it with the existing AGIMM algorithm can improve the convergence speed of turning rate and enhance tracking performance.

[0045] 3.1 Target Maneuver Identification

[0046] Target maneuver discrimination is primarily based on residual information from model predictions. After the k-th filtering iteration, the three models of the AGIMM algorithm will each produce three residual vectors. These correspond to three posterior probability values. Using the residual of model i with the largest posterior probability value as the criterion, calculate its distance function:

[0047] In the formula: Let be the residual of model i at time k; Let be the residual covariance of model i at time k.

[0048] It obeys the measurement dimension value. Distribution. (Through query) The distribution table allows setting a threshold value M for maneuvering turns. If... If a maneuver is detected, it is determined that a maneuver has occurred; otherwise, it is determined that no maneuver has occurred. In the case of a maneuver, the discrete turning rate is recalculated based on historical data points, updating the original turning rate and incorporating it into the posterior probability calculation and ranking, guiding the convergence of the turning rate calculation in the AGIMM algorithm. If no maneuver is detected, the turning rate of the maneuver discrimination model is maintained, and it also participates in the posterior probability calculation and ranking.

[0049] 3.2 Calculation of Discrete Turning Rate

[0050] The curvature at a point can reflect the turning rate at that point. Therefore, the accuracy of tracking the turning rate of a target can be improved by calculating discrete curvature. This invention utilizes the method of calculating discrete curvature to assist the AGIMM algorithm in converging the turning rate. By extracting some historical estimated points, the influence of noise can be reduced, and the curvature of the trajectory can be calculated. Historical points need to be extracted at intervals, such as... Figure 1 As shown.

[0051] The formula for calculating the discrete curvature at point B is:

[0052] In the formula: Let B be the discrete curvature. These are the lengths of line segments AB, BC, and AC, respectively. Let be the directed area of ​​triangle ABC. The formula for calculating the directed area is:

[0053] In the formula: i, j, and k are all unit vectors. Using the calculated discrete curvature, the target turning rate can be further calculated, as shown in the formula: (10)

[0054] 3.3 Adjustment of the Model Structure Based on Ranking

[0055] Turn rate calculation methods based on discrete curvature require the use of historical estimated track points. Therefore, estimations of maneuvers with continuously changing turn rates can be significantly biased due to the influence of these historical track points. However, turn rate estimation methods based on maneuver discrimination can promptly compensate for the shortcomings of discrete curvature-based methods. The IMM model can effectively achieve this complementarity between models. To address this, this invention proposes the MDS-AGIMM algorithm, which, through the introduction of a sorting mechanism, better facilitates interaction between models.

[0056] After the k-th filtering is completed, the probability value of the turning rate is obtained based on the AGIMM model and the discrete curvature calculation model, respectively. The first three terms are the posterior probabilities of the turning rate obtained from AGIMM, and the fourth term is the posterior probability obtained from the discrete curvature calculation model. The convergence direction of the turning rate is determined based on the magnitude of the turning rate, and there are four cases in total.

[0057] (1)

[0058] At this point, the model error for calculating the discrete curvature is relatively small, so the turning rate calculated by this model should be prioritized to guide the AGIMM model towards iterative convergence of the turning rate. The turning rates of the four models are then sorted from largest to smallest based on their posterior probability values, denoted as . Its corresponding turning rate

[0059] like ,but:

[0060] In the formula: α is an adjustment factor that affects the convergence speed of the turning rate, and is usually taken as 1 / 2; This is the minimum grid spacing. If If no maneuver occurs, the turning rate is updated as follows:

[0061] In the formula, This is the threshold value for an invalid model, indicating that the search error for the turning rate in this direction is large, and the model should be moved closer to the grid center. At this point:

[0062] At this point, the left-jump model has the optimal probability, and we first sort them according to their probability values. If If a maneuver is detected, the model turning rate is updated according to equation (12), and the model spacing adjustment parameter becomes:

[0063] like If no maneuver occurs, the center model turning rate update method is calculated according to equation (14), and the left jump and right jump model update methods are as follows:

[0064] At this time, the grid adjustment method is the same as that in equation (17).

[0065]

[0066] At this point, the intermediate model has the optimal probability, and we first sort them according to their probability values. If If the maneuver is detected, the turning rate adjustment is the same as in (12), and the model spacing adjustment parameter is:

[0067] like If no maneuver occurs, the turning rate adjustment method and grid adjustment method shall be calculated according to formulas (14) to (17).

[0068]

[0069] At this point, the right-jump model has the optimal probability, and we first sort them according to their probability values. If If a maneuver is detected, the model turning rate update method is calculated according to equation (12), and the model spacing adjustment parameter becomes:

[0070] like If no maneuver occurs, the center model turning rate update method is calculated according to equation (14), and the left jump and right jump model update methods are as follows:

[0071] At this time, the grid adjustment method is the same as that in equation (17).

[0072] 3.4 Adjustment of the model probability transition matrix

[0073] The probability transition matrix is ​​also adaptively adjusted based on the residual information after each filtering step. The smaller the residual, the higher the probability of the model occurring, and other models have a higher probability of transitioning to that model. When the target maneuvers and the residual value changes, the probability transition matrix is ​​adjusted accordingly, increasing the probability of the model set transitioning to the model with the smaller residual. The update method is as follows:

[0074] In the formula: Let be the likelihood function of model j after filtering at time k; Let be the transition probability from model i to model j at time k. However, adjusting the model transition probability in this way still has shortcomings. When a certain model probability value is too large or too small, it will make it difficult for the model to jump out and unable to transition in time. Therefore, further improvement is needed to limit the maximum and minimum values ​​of the probability.

[0075]

[0076] 4. Simulation Analysis

[0077] 4.1 Target movement trajectory

[0078] Track 1: Geological Exploration Track

[0079] The areas surrounding the airport are mostly suburban, flat, and underdeveloped plains or fields. There is a demand for small to medium-sized UAVs used for agricultural and forestry operations or geological exploration. These targets are analyzed as tracking targets. The UAV's initial position is [0m, 0m, 0m], and its initial speed is [0m / s, 5m / s, 0.5m / s]. The target flight is divided into seven stages: Stage 1 (0-100 seconds): The UAV flies at a constant initial speed; Stage 2 (100-118 seconds): The target turns at a turn rate of -5° in the horizontal plane; Stage 3 (118-218 seconds): The target continues to fly at a constant speed; Stage 4 (218-254 seconds): The target turns at a turn rate of 5° in the horizontal plane; Stage 5 (254-354 seconds): The target flies at a constant speed in a straight line; Stage 6 (354-390 seconds): The target turns at a turn rate of -5° in the horizontal plane; Stage 7 (390-490 seconds): The target flies at a constant speed in a straight line. Flight track Figure 2 As shown.

[0080] Track 2: Drone aerial photography trajectory

[0081] In some consumer drone groups, action photography flights account for a large proportion. This demand exists not only among the general public but also among some government agencies and businesses. However, due to a lack of public awareness of flight control regulations, there is a possibility of unauthorized flights over important targets, necessitating the tracking and monitoring of such illegal flights. These drones, due to their relatively simple equipment and limited operator skills, exhibit free and unpredictable flight paths with varying turning maneuvers. To address this, flight path two is designed with the same initial position and state as path one, dividing the flight into two phases. The first phase, from 0 to 100 seconds, involves the drone flying at a constant initial speed in a straight line. The second phase, from 100 to 400 seconds, involves the drone initially turning at a 5° angle and then spiraling upwards according to a varying turning rate. The formula for the changing turning rate is:

[0082]

[0083] Flight track Figure 3 As shown.

[0084] 4.2 Performance Evaluation Indicators

[0085] The root mean square error (RMSE) was used as the metric for comparing the performance of the methods. The calculation formula is as follows:

[0086] In the formula: Let i be the value of the i-th component of the state vector at time k. This is the estimated value of the i-th component of the state vector at time k.

[0087] 4.3 Comparative Analysis

[0088] Target tracking simulation was performed on the constructed virtual flight scenario. The parameter settings of the model are shown in Table 1.

[0089] Table 1 Model Parameter Settings

[0090] By comparing the tracking results of angular velocities, the performance advantages of the method can be demonstrated, such as... Figure 4 As shown;

[0091] The comparison of simulation curves shows that the improved AGIMM algorithm of this invention has better accuracy and robustness in tracking angular velocity than the original method. This is mainly because the discrete curvature calculation is performed through point-by-point sampling, reducing the impact of disturbances on the turning rate. On the other hand, by sorting the residual information, poor-quality models are eliminated, further reducing the impact of disturbances. The changes in the trajectory-REMS values ​​are shown in Figure 5.

[0092] From the perspective of target path tracking performance in various directions, the method designed in this invention achieves higher accuracy than the AGIMM algorithm. Table 2 shows the REMS values ​​obtained from 100 Monte Carlo simulations of the model, categorized by direction.

[0093] Table 2 Comparison of REMS values

[0094] As can be seen from Table 2, the improvement in tracking accuracy is mainly reflected in the X and Y directions, while the improvement in the Z direction is almost the same. This is because the improvement of the method in this invention is mainly aimed at enhancing the adaptive tracking capability of the turning rate of the horizontal turning model, and no improvement has been made in the vertical direction.

[0095] Curve 2 focuses on the tracking performance of the method for continuously changing angular velocities. The angular velocity tracking performance is as follows: Figure 6 As shown;

[0096] from Figure 6 As can be seen, for continuously changing angular velocities, the improved AGIMM algorithm still exhibits better tracking accuracy and robustness compared to the original method. The comparison of REMS values ​​is similar to that of trajectory one, and will not be repeated here.

[0097] 5. Conclusion

[0098] Tracking and monitoring low-altitude, slow-moving, and small targets in horizontal maneuvers has always been a challenge in low-altitude airspace management. Traditional multi-model algorithms can significantly improve target tracking accuracy through interaction between models, but their accuracy is limited for targets undergoing continuous maneuvers because the target's maneuverability cannot be known in advance. The AGIMM algorithm addresses the adaptive tracking problem of turn rate, but its tracking accuracy is affected by disturbances. To address this, this invention incorporates a discrete turn rate calculation model and, through sorting, promptly removes models with large residuals. This process enhances the model's robustness. Simulations of typical operational paths of small and medium-sized UAVs demonstrate that the algorithm has good tracking capabilities for both intermittent and continuous turn maneuvers, especially in terms of turn rate tracking, exhibiting better robustness compared to the traditional AGIMM model.

Claims

1. An AGIMM tracking method based on motion detection and sorting, characterized in that: Includes the following steps: S1. Establish a cooperative turning tracking model; S2. Update the turning rate of the cooperative turning tracking model using the AGIMM algorithm; S3. Identify target maneuvers; S4. Calculate the discrete turning rate using the discrete turning rate calculation model; S5. Adjust the model structure; S6. Adjust the model probability transition matrix; In step S5, adjusting the model structure includes the following steps: S51, when At this point, the turning rate calculated by the discrete turning rate calculation model is used as the main factor. The turning rates are sorted from largest to smallest according to the various posterior probability values, which are as follows: , , , Its corresponding turning rate , , , ;like ,but: ； in, , , , The first three terms are the posterior probabilities of the turning rate obtained from AGIMM, and the fourth term is the posterior probability obtained from the discrete curvature calculation model. The turning rate of the left jump model. For the intermediate model's turning rate, The turning rate of the right jump model; In the formula: As a regulating factor; Minimum grid spacing; like If no maneuver occurs, the turning rate update formula is: ； ； In the formula, This is the threshold value for an invalid model; at this point: S52, when At this point, the turning rate of the left-jump model is the primary factor, and the models are sorted according to their posterior probability values; if The system is determined to have engaged in maneuvering; the model spacing adjustment parameters are as follows: like If no maneuver occurs, the turning rate update formula is: in, This is the threshold value for maneuvering and turning. It is a distance function; S53, when At this point, the intermediate model's turning rate is the primary factor, and the models are sorted based on their posterior probability values; if If a maneuver occurs, the model spacing adjustment parameter is determined to be: like No maneuver occurred; S54, when At this point, the right-jump model's turning rate is the primary factor, and the models are sorted based on their posterior probability values; if The system is determined to have engaged in maneuvering; the model spacing adjustment parameters are as follows: like If no maneuver occurs, the turning rate update formula is:

2. The AGIMM tracking method based on motion detection and sorting as described in claim 1, characterized in that: In step S1, assume the UAV performs a turning maneuver in the horizontal plane, with the velocity scalar remaining constant during the turn and the angular velocity expressed as... Therefore, the calculation formula for the cooperative turning tracking model is: In the formula: The spatial position and velocity of the target; The system is a discrete-time white noise sequence; ； The covariance matrix of the noise is: ； In the formula: when ω > 0, the target turns left; when ω < 0, it turns right; when ω = 0, it becomes a uniform motion model; q is the system noise variance.

3. The AGIMM tracking method based on motion detection and sorting as described in claim 2, characterized in that: In step S2, the turning rates of the three cooperative turning tracking models are arranged from smallest to largest, specifically as follows: ,Will As the turning rate of the left jump model. As an intermediate model turning rate As the turning rate of the right jump model; the posterior probability corresponding to each turning rate is: In each After the filtering of the wheels is completed, each turning rate is updated according to the corresponding posterior probability.

4. The AGIMM tracking method based on motion detection and sorting as described in claim 3, characterized in that: In step S3, the determination of the target maneuver includes the following steps: S31. After the k-th filtering, the three cooperative turning tracking models updated by the AGIMM algorithm yield three residual vectors respectively. These correspond to three posterior probability values. Using the residual of the cooperative turning tracking model i with the largest posterior probability value as the criterion, calculate its distance function, which is given by the following formula: In the formula: The residual of the cooperative turning tracking model i at time k; Let i be the residual covariance of the collaborative turning tracking model i at time k; S32. Set the threshold value M for maneuvering and turning; if If the condition is met, it is determined that a maneuver has occurred; otherwise, it is determined that no maneuver has occurred.

5. The AGIMM tracking method based on motion detection and sorting as described in claim 4, characterized in that: In step S4, the calculation formula for the discrete turning rate calculation model is as follows: ； In the formula: For discrete turning rates; These are the lengths of the line segments between discrete points A, B, and C, respectively. Let be the directed area of ​​triangle ABC.

6. The AGIMM tracking method based on motion detection and sorting as described in claim 1, characterized in that: In step S6, the formula for calculating the model probability transition matrix is: In the formula: Let be the likelihood function of model j after filtering at time k; Let i be the likelihood function of model i at time k after filtering; Let be the transition probability from model i to model j at time k; After improvement:

Images