Maneuvering extended target tracking method based on golden section adaptive grid

Abstract

The invention relates to a maneuvering extended target tracking method based on a golden section self-adaptive grid, which comprises the following steps of: firstly, representing a model set for tracking in a grid form, and dividing the grid into an internal grid and a peripheral grid; then generating a local weight and an overall weight of the model set, arranging the local weight and the overall weight with the two models in the internal grid according to a golden section proportion, and calculating a distance between the local weight and the overall weight to reflect a difference between the local weight and the overall weight and a target real motion mode matching degree so as to serve as a basis for model parameter updating; and finally, according to a result obtained by filtering the updated model set, updating a transition probability matrix, and according to a model filtering residual error with the maximum model probability, judging the maneuvering degree of a target so as to determine whether each module of the algorithm is executed or not, thereby improving the efficiency of the algorithm. The method provided by the invention can effectively solve the problem that the tracking precision of an extended target tracking method in a strong maneuvering scene is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of maneuvering extended target tracking, and particularly to a maneuvering extended target tracking method based on golden-section adaptive grid. Background Art

[0002] Radar target tracking technology has made remarkable progress in the past few decades. Due to the limited resolution of previous radar sensors, many tracking algorithms are based on a single scattering center (i.e., the measurement source) obtained from the object surface by the sensor at each moment, regarding the object as a point target and estimating its motion state (i.e., position, velocity, and acceleration). With the improvement of the resolution of modern sensors, multiple measurement sources can be resolved from the object surface in each scan, and thus the target is called an extended target. Modern air combat usually involves high-speed and highly maneuverable aviation weapons. Tracking these objects and conducting precise strikes require high-precision tracking algorithms. Therefore, the maneuvering extended target tracking technology has been developed.

[0003] Maneuvering extended target tracking is a process of jointly estimating the motion state and shape information of a target based on the measurement information obtained from a sensor. Due to the strong coupling relationship between the motion state and shape information of the target, accurate estimation of the target's motion state is beneficial to obtaining a more accurate target shape, while an accurate target shape also helps in the accurate estimation of the target's motion state. Any target tracking algorithm is based on a motion model. Therefore, a good motion model is crucial for the performance of a maneuvering extended target tracking algorithm.

[0004] Existing single motion models include the Constant Velocity (CV) model, the Constant Acceleration (CA) model, the Constant Turn (CT) model, the current statistical model, and the Jerk model, etc. However, relying solely on a single motion model is insufficient to accurately describe the motion state of an object, especially when the object has complex and variable motion behaviors. When the target suddenly moves, the limitations of a single motion model may even lead to tracking failure. To address this challenge, Blom and Bar-Shalom first proposed the Interactive Multiple-Model (IMM) algorithm. The IMM algorithm uses a fixed set of multiple motion models and performs parallel filtering at each time step. The filtering results are weighted and fused based on the corresponding model probabilities, and the fused result represents the estimated result of the state. The probability of each model changes over time. However, the IMM algorithm relies on a predefined set of models with fixed parameters, and it can achieve the best performance only when the pre-designed models match the true motion pattern of the target. To consider a wider range of potential motion patterns, the number of models is usually increased. However, this approach inevitably increases the computational complexity and introduces competition between models, ultimately reducing the tracking performance.

[0005] To overcome the fixed drawbacks of the IMM model, Li and Bar-Shalom proposed the Variable-Structure Multiple-Model (VSMM) algorithm. The difference between the VSMM algorithm and IMM lies in that the motion models used in the VSMM algorithm are variable, including changes in quantity and model parameters, so as to be able to adaptively select appropriate motion models according to the external environment. Both the IMM and VSMM algorithms need to perform state estimation on the target based on a set of models. Therefore, the design method of the model set is crucial for the performance of the VSMM algorithm. The existing model set design methods are mainly divided into two categories: (1) Preset the parameters of many motion models in advance, and then activate the motion models among them as a new model set according to different situations. (2) Generate new motion models in real time according to the results of state estimation to update the model set. Among the first category of algorithms, the common ones are the Model Group Switch (MGS) algorithm and the Likely Model Set (LMS) algorithm, but both are highly dependent on the preset and division of the model set structure and cannot activate models outside the model set. And the Expected Model Augmentation (EMA) algorithm, as the second category of model set design method, can perform weighted fusion on the models according to the filtering results at each moment to generate a new expected model and expand it into the model set. The EMA algorithm utilizes real-time information and has higher performance compared with the IMM algorithm.

[0006] However, due to the strong coupling relationship between the motion state and shape of the extended target, a good motion state estimation result is beneficial to the accurate description of the extended shape. In order to enable the EMA algorithm to achieve better tracking performance in the tracking of maneuvering extended targets, it still has the following limitations to overcome: (1) The basic model of the EMA algorithm consists of a set of motion models with fixed parameters, and the transition probabilities between the models are also fixed. Therefore, the ability of the expected model generated based on the basic model to modify the model set and the efficiency of model switching are limited. (2) The generated expected model represents the weighted fusion result of the entire model set and cannot provide a detailed description of the local structure of the model set. (3) The EMA algorithm updates the expected model at each moment. If the generated expected model matches the real pattern very well, continuing to update it will bring unnecessary computational burden.

[0007] The technical problem that needs to be urgently solved by those skilled in the art currently is: how to improve the existing EMA model set design method to make it achieve the adaptability of the structural parameters of the model set and the transition probabilities between the models, so as to further improve the tracking accuracy of extended targets in a strong maneuvering scenario. Summary of the Invention

[0008] To address the deficiencies in the prior art, the present invention provides a maneuvering extended target tracking method based on the golden-section self-adaptive grid (i.e., GSAG-EMA), which can achieve the self-adaptation of the structural parameters of the model set and the transition probabilities between models, and improve the tracking accuracy and algorithm execution efficiency of extended targets in strong maneuvering scenarios.

[0009] To achieve the above object, the specific solution adopted by the present invention is as follows: A maneuvering extended target tracking method based on the golden-section self-adaptive grid, comprising the following steps:

[0010] S1. To enable the models in the model set to be reasonably distributed, thereby ensuring the accuracy of the model set approximating the true pattern, the model set is represented in the form of a grid and divided into an internal grid and an external grid. The models are represented as points in the grid, and the parameters of each model are initialized.

[0011] S2. Using the model probabilities corresponding to each model in the model set, the model parameters are weighted to generate an overall weighted value, i.e., the expected model.

[0012] S3. Calculate the weighted value of the internal grid models to generate a local weight. Arrange the overall weight, the local weight, and the two models whose initial positions in the internal grid are on the left and right sides of the spatial center point according to the golden-section ratio, and adaptively adjust the parameters of the models in the internal grid and the external grid, thereby generating an effective model set.

[0013] S4. According to the model set obtained in step S3, perform IMM estimation on each model, and update the probability of each model and the transition probability matrix (Transition Probability Matrix, TPM).

[0014] S5. According to the magnitude of the model filtering residuals corresponding to the maximum model probability, divide the maneuvering intensity of the target into weak maneuvering, medium maneuvering, and strong maneuvering, and adopt different algorithm execution strategies at different maneuvering levels.

[0015] S6. Take the adjusted model parameters and the updated transition probability matrix as the input of the model conditional filter, and repeat steps S2 - S5.

[0016] Further, in S1, the CA model and the CT model are selected as the research objects. Since the dimensions of the CA model and the CT model corresponding to uniform acceleration motion and uniform rate turning motion are different, the grid model division and the initial definition of the model parameters for the two are as follows:

[0017] (1) For the CA model, its internal and external grid models are initialized as:

[0018]

[0019]

[0020]

[0021] Among them, A c represents the model space of the CA model and is a two-dimensional space, and a max represents the maximum acceleration value that the target can reach, and a max = 40m / s 2 ; represents the external grid model, and respectively represent the upper edge, lower edge, left edge and right edge models of the initial external grid model; represents the internal grid model, and respectively represent the internal grid models above, below, to the left and to the right of the center model of the internal grid at the initial position.

[0022] (2) For the CT model, its internal and external grid models are initialized as:

[0023] W c = {w: -w max ≤ w ≤ w max}

[0024]

[0025]

[0026] Among them, W c represents the model space of the CT model and is a one-dimensional space, and w max represents the maximum angular velocity that the target can reach, and w max = 10rad / s; represents the internal grid model, and respectively represent the internal grid models to the left and to the right of the center model ; represents the external grid model, and respectively represent the left edge and right edge models of the initial external grid model.

[0027] Furthermore, in S2, according to the model probability, the generation rule of the expected model is: Among them, M k-1 represents the set of valid models at the (k - 1)th moment, is the jth basic model The corresponding probability.

[0028] Furthermore, in S3, first, the internal grid model is weighted according to the model probability to obtain the local weighting of the model set; then, the overall weighting, the local weighting, and the two models with the initial positions of the internal grid on the left and right sides of the spatial center point are spatially arranged according to the golden ratio. The distance between the two weightings is used to reflect the difference in their matching degrees to the true motion pattern, so as to analyze the maneuverability of the target and adaptively adjust the parameters of the models in the internal grid and the external grid to generate an effective model set.

[0029] (1) For the CA model, the adaptive rules for its internal and external grid models are as follows:

[0030] ① Adaptive rule for the internal grid model

[0031]

[0032]

[0033]

[0034]

[0035] Where is the central model of the internal grid and also the local weighting of the model set; dist represents the Euclidean distance between the local weighting and the overall weighting, which is used to reflect the difference in their matching degrees to the true motion pattern. When the Euclidean distance between the two changes, the shape structure of the internal grid model space will change accordingly; represents the internal grid model at time k, where and are centrosymmetric about so that is located at the golden section point of the line segment between and At the same time, it also exactly makes be located at the golden section point between and Define the distance between and as distance; and are the models with the vector formed by to as the positive x-axis direction and are located in the upper left, upper right, lower left, and lower right of respectively. p is a proportionality coefficient used to adjust the positions of and .

[0036] ② External grid model adaptation rules

[0037]

[0038]

[0039]

[0040]

[0041] Among them, Distance represents the distance to the origin, and ψ represents the angle formed by the origin and the positive x-axis; S1, S2, S3, S4, and S5 represent the respective regions after dividing the model space of the CA model; p n (n = 1, 2, …, 5) represents the proportionality coefficient; and respectively represent the upper edge, lower edge, left edge, and right edge models of the external grid model at time k.

[0042] (2) For the CT model, the adaptation rules for its internal and external grid models are as follows:

[0043] ① Internal grid model adaptation rules

[0044]

[0045]

[0046]

[0047] Among them, is the central model of the internal grid and is also the local weighting of the model set. and are symmetric about the center, and is at the golden section point of the line connecting to is at the golden section point of the line connecting to and are the internal grid models of the CT model at time k.

[0048] ② External grid model adaptation rules

[0049]

[0050]

[0051]

[0052] Among them, Distance1 represents the distance from the expected model of the CT model to the origin; the model space of the CT model is divided into three segments, namely L1, L2, and L3; and represents the external grid model at time k.

[0053] (3) The generation rule of the effective model set is as follows:

[0054]

[0055] Among them, M k represents the effective model set at time k, and represent the expected model, the internal grid model, and the external grid model at time k, respectively.

[0056] Furthermore, in S4, the effective model set generated in step S3 is used to perform IMM estimation on each model to obtain the updated model probability, and then the transition probability matrix is adaptively updated:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] Among them, is the probability of model j updated at time k, is the correction factor of the element in the i-th row and j-th column of the transition probability matrix at time k, λ is the correction rate, is the updated transition probability matrix at time k, and r represents the number of models in the model set.

[0064] Furthermore, S5 uses the filtering residual of the model corresponding to the maximum model probability in S4 as the basis for judging the maneuverability of the target, and decides whether to execute each module of the proposed method: For model j, its filtering residual is: Assume that the maximum model probability value at time k is: Assume that the filtering residual of the model corresponding to the maximum model probability is D k , under different maneuver levels, the implementation of the proposed method is as follows: Among them, is the innovation of model j, is the covariance of the innovation, T represents the transpose symbol, and D1 and D2 are two thresholds used to divide the target maneuver level according to the magnitude of the filtering residual value

[0065] Beneficial effects: Under the framework of improving the maneuvering extended target tracking of the variable structure multiple model, through the model set design method of augmenting the adaptive grid expectation model based on the golden section, the model set is divided into two parts: the internal grid and the external grid. The local weighting and overall weighting of the model set, as well as the two models whose initial positions in the internal grid are on the left and right sides of the spatial center point, are arranged in space according to the golden section ratio, and the model parameters are updated according to the difference between the local and the overall as the judgment basis, so as to improve the situation that the tracking algorithm is prone to accuracy decline in complex and changeable maneuvering scenarios, enabling the tracking algorithm to more accurately estimate the motion state and morphological information of the target, improving the tracking accuracy. Then, an adaptive update strategy of the transition probability matrix (TPM) is added in the process of IMM estimation, so as to improve the switching efficiency between models; finally, according to the filtering residual of the model with the maximum model probability, the maneuver degree of the target is divided, and whether each module of the tracking algorithm is executed is determined according to different maneuver degrees, thereby further improving the execution efficiency of the algorithm. Description of the Drawings

[0066] Figure 1 is a schematic diagram of the internal grid structure of the CA grid;

[0067] Figure 2 is a schematic diagram of the spatial region division of the CA model;

[0068] Figure 3 is a schematic diagram of the external grid structure of the CA grid;

[0069] Figure 4 is the effect diagram of the motion trajectory and morphological estimation in the CA grid simulation experiment;

[0070] Figure 5 is the comparison diagram of the Hausdorff distance and root mean square error of four algorithms in the CA grid simulation experiment;

[0071] Figure 6 is the schematic diagram of the internal and external grid structures of the CT grid;

[0072] Figure 7 It is the effect diagram of motion trajectory and shape estimation in the CT grid simulation experiment;

[0073] Figure 8 It is the comparison diagram of Hausdorff distance and root mean square error of four algorithms in the CT grid simulation experiment;

[0074] Note: The root mean square error and Hausdorff distance are respectively used to evaluate the motion state of the target and the estimation accuracy of the shape. The smaller their values are, the higher the accuracy of the algorithm is. Specific implementation manners

[0075] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0076] A maneuvering extended target tracking method based on the golden section adaptive grid includes steps S1 to S6.

[0077] Step S1: Grid division and parameter initialization of the model set: The model set that maximally covers the motion space of the target and accurately describes the true motion mode of the target is represented in the form of a grid, and the grid is divided into an internal grid and an external grid. Any point in the grid corresponds to a modality, and the parameters of each model are initialized.

[0078] Any point in the model space corresponds to a modality. We need several modalities to form a model set to approximately describe the true motion mode of the target. However, these modalities cannot be too scattered or concentrated, otherwise it will lead to a reduction in the accuracy of the model set approximating the true mode. In order to make the models in the model set be reasonably distributed, so as to ensure the accuracy of the model set approximating the true mode, we need to perform spatial division on the model set, so that the models in the model set are reasonably distributed, represent the model set in the form of a grid, and divide it into an internal grid and an external grid.

[0079] The design of the internal grid is to meet the requirement that the structure and parameters of the model set can change efficiently under complex and changeable maneuvering scenarios, so as to have better adaptability. The models of the external grid are mainly located at the edge of the model space, and the change amplitude is much smaller than that of the internal grid. This is to ensure a certain degree of stability while having adaptability. The model structures and parameters in the internal grid and the external grid change with the update of the expected model, and jointly form a complete model set.

[0080] The motion of most targets can be decomposed into linear acceleration (CA) and turning motion (CT). The CA model and CT model can better describe these basic motion patterns, and the CA and CT models have simple structures, small computational amounts, and can be extended to other models, and are widely used in fields such as radar and drones. Therefore, the present invention selects the CA model and CT model as the research objects; since the dimensions of the CA model and CT model are different, the CA model is two-dimensional and the CT model is one-dimensional, so the grid models for the CA and CT models are designed separately. The grid model division and the initial definition of the model parameters (at the initial moment when k = 0, set the initial values of the parameters of the grid model) are as follows

[0081] (1) Parameter initialization of the CA grid model

[0082] The model space A of the CA model c is a two-dimensional space, and any point in the space corresponds to a mode:

[0083]

[0084] The model set of the CA model contains 9 models. In order to make the model set of the CA model cover the true motion pattern of the target as reasonably as possible, the initial value of the external grid model is set to:

[0085]

[0086] Due to the two-dimensional characteristics of the CA model, if the structure and parameters of the internal grid want to achieve better adaptability, the initial positions of the five models of the internal grid should be in a cross-shaped distribution, and their initial values should not be too large or too small. Being too large will cause duplication with the external grid model, and being too small will cause the models to pile up together and unable to perform effective adaptive structure and parameter updates; therefore, in order to make the internal grid have good self-adaptability when facing changes in the target maneuverability, the initial value of the internal grid model is set to:

[0087]

[0088] Among them, A c represents the model space of the CA model and is a two-dimensional space, a max represents the maximum acceleration value that the target can reach, a max = 40m / s 2 ; represents the external grid model, and respectively represent the upper edge, lower edge, left edge and right edge models of the initial external grid model; represents the internal grid model, and respectively represent the internal grid models above, below, to the left, and to the right of the central model at the initial position in the internal grid.

[0089] (2) Parameter initialization of the CT grid model

[0090] Due to the one-dimensional nature of the CT model, the internal grid is mainly responsible for the adaptability of the model set, and the external grid needs to ensure the stability of the model set during changes. Therefore, the initial parameter values of the CT model are set as follows:

[0091] W c = {w: -w max ≤ w ≤ w max}

[0092]

[0093]

[0094] where W c represents the model space of the CT model and is a one-dimensional space, w max represents the maximum angular velocity that the target can reach, w max = 10 rad / s; represents the internal grid model, and respectively represent the internal grid models to the left and to the right of the central model ; represents the external grid model, and respectively represent the left edge and right edge models of the initial external grid model.

[0095] S2: Augment an expected model in the original model set. This model is obtained by probabilistically weighted summation of a group of j basic models, aiming to match the expected value of the target true motion pattern.

[0096] where M k-1 represents the effective model set at the (k - 1)th moment, is the probability corresponding to the jth basic model .

[0097] S3: Calculate the weighted value of the internal grid model to generate local weights, thereby reflecting the local situation of the model set; then arrange the overall weight, local weight, and the two models in the internal grid whose initial positions are on the left and right sides of the spatial center point according to the golden ratio. Reflect the difference in their matching degrees to the true motion pattern based on the distance between the two weights, so as to analyze the maneuverability of the target and adaptively adjust the parameters of the models in the internal grid and external grid to generate an effective model set.

[0098] (1) Parameter Adaptation of CA Grid Model

[0099] ① Parameter Adaptation Rule of Internal Grid Model

[0100] If the target makes maneuvers of different degrees, the shape and size of the internal grid should also change accordingly to achieve the self - adaptability of the internal grid model. Obviously, the local weight and the overall weight have different matching degrees according to different motion patterns of the target, which can be reflected by to the Euclidean distance between them. When the Euclidean distance between the two changes, the shape structure of the internal grid model space will change accordingly.

[0101]

[0102]

[0103]

[0104]

[0105] distance = 1.618 * dist

[0106]

[0107]

[0108]

[0109]

[0110] Rot θ (a, b) = (acosθ - bsinθ, asinθ + bcosθ)

[0111]

[0112]

[0113]

[0114]

[0115] Reference Figure 1 , where is the central model of the internal grid and also the local weighting of the model set; dist represents the Euclidean distance between the local weighting and the overall weighting, which is used to reflect the difference in the matching degree between the two and the true motion pattern; represents the internal grid model at time k, and respectively represent two models in the internal grid generated by the CA model according to the golden ratio, and are centrosymmetric about so that is located at the golden section point of the line segment between and ; at the same time, it also exactly makes be located at the golden section point between and . Define the distance between and as distance; and respectively take the vector formed by and as the positive x-axis direction and are located in the upper left, upper right, lower left, and lower right of ; p is the proportionality coefficient used to adjust the positions of and .

[0116] ② Parameter Adaptive Rule of External Grid Model

[0117] Since the expected model will be in different positions in the model space over time, it is considered to adjust the parameters of the external grid model according to the position where is located, change the model space of the external grid, so that the overall grid is more inclined to match the true motion pattern of the target. The model space of the CA model in the present invention is divided into five parts, namely S1, S2, S3, S4, and S5, as shown in Figure 2 .

[0118] Let the distance from the expected model to the origin be:

[0119]

[0120] Then the regional division rules of S1, S2, S3, S4, and S5 are:

[0121]

[0122] The specific calculation rules for the external grid model parameters are as follows:

[0123]

[0124]

[0125] Among them, Distance represents the distance to the origin, and ψ represents the angle formed with the positive x-axis direction to the origin; referring to Figure 3 , when are respectively in S1, S2, S3, S4, and S5, the corresponding external grid model makes different changes; p n (n = 1, 2, …, 5) represents the corresponding proportionality coefficient when in different regions, and respectively represent the upper edge, lower edge, left edge, and right edge models of the external grid model at the kth moment.

[0126] (2) Parameter adaptation of the CT grid model

[0127] ① Parameter adaptation rules for the internal grid model

[0128]

[0129]

[0130]

[0131] Referring to Figure 6 , in the CT model, the two models of the internal grid refer to and Among them, and are centrosymmetric about , and is at to the golden section point of the connection line, is at to the golden section point of the connection line. and are the internal grid models of the CT model at the kth moment.

[0132] ② Parameter adaptation rules for the external grid model

[0133]

[0134]

[0135]

[0136] Reference Figure 6 , where Distance1 represents the distance from the expected model of the CT model to the origin; the model space of the CT model is divided into three segments, namely L1, L2, and L3; and represents the external grid model at time k.

[0137] (3) Expected model is designed to match the expected value of the target's true motion pattern. However, when the maneuver level of the target changes significantly, if the model parameters of the basic model set remain fixed, it will lead to a decrease in the matching degree of the expected model to the true motion pattern, and even cause a decrease in the tracking accuracy of the algorithm. Therefore, the present invention uses an expected model augmentation algorithm to obtain the expected model under the variable basic model set to achieve the adaptability of the effective model set M k ; the generation rule of the effective model set is:

[0138]

[0139] where M k represents the effective model set at time k, and represent the expected model, the internal grid model, and the external grid model at time k, respectively.

[0140] S4: After performing IMM estimation on each model in the effective model set obtained in step S3, update the probability of each model and the transition probability matrix: where, is the probability of model j updated at time k, is the correction factor of the element in the i-th row and j-th column of the transition probability matrix at time k, λ is the correction rate, is the updated transition probability matrix at time k, and r represents the number of models in the model set.

[0141] S5: Since the target is not always performing high-intensity maneuvers, updating the transition probability matrix and adjusting the model parameters of the grid model at any time will undoubtedly increase the computational burden, thereby reducing the speed of the algorithm. Therefore, the present invention classifies the maneuver intensity of the target into weak maneuver, medium maneuver, and strong maneuver according to the size of the model filtering residual corresponding to the maximum model probability, and adopts different algorithm execution strategies at different maneuver levels to further improve the execution efficiency of the algorithm.

[0142] For model j, its filtering residual is as follows:

[0143] where is the innovation of model j, is the covariance of the innovation, and T represents the transpose symbol.

[0144] Assume that the maximum model probability value at time k is:

[0145]

[0146] Assume that the filtering residual of the model corresponding to the maximum model probability is D k , and under different maneuver levels, the execution of the proposed method is as follows:

[0147] where D1 and D2 are two thresholds used to divide the target maneuver level according to the magnitude of the filtering residual value.

[0148] That is, when the maneuver level of the target is low, only IMM estimation needs to be performed, and there is no need to update the model parameters and the transition probability matrix; when the maneuver degree of the target is at a medium level, while performing IMM estimation, GSAG-EMA is executed, and the transition probability matrix is not updated; when the maneuver level of the target is high, IMM estimation, GSAG-EMA, and the update of the transition probability matrix need to be performed simultaneously; through this hierarchical execution, compared with the case of executing all parts of the algorithm at any time, the efficiency of the algorithm can be improved.

[0149] S6: Use the adjusted model parameters and the updated transition probability matrix as the input of the model conditional filter, and repeat steps S2 - S5.

[0150] The innovations of the present invention are reflected in: (1) Representing the model set in the form of a grid, dividing the grid into internal and external regions, and the model parameters within the grid are updated in real time according to specific rules instead of being preset. (2) Performing weighted fusion on the internal grid models and the entire grid model respectively to generate local and overall weighted results for the model set, arranging these results, as well as two models from the internal grid, in the ratio of the golden section, calculating the distance between the local and overall weighted results to reflect their matching degree with the true pattern, and providing detailed rules for grid structure change and model parameter update. (3) Updating the transition probability between models based on the filtering results, calculating the filtering residual of the model with the highest probability, using it to evaluate the maneuver degree of the object, and deciding whether to execute each step of the algorithm based on this criterion, thereby improving the calculation efficiency.

[0151] The effectiveness of the present invention is verified through simulation experiments. The simulation experiments verify the CA and CT grid models respectively. A single star-convex maneuvering extended object is used as the experimental tracking object, and the Monte Carlo method is used for the experiments, without involving clutter and missed detection problems. The program is executed 100 steps at a time, and the sampling time is set to 1 s. In the experiment for verifying the CA grid model, the maneuvering process of the target is set as follows: it moves at a constant speed from 0 to 25 s, accelerates with an acceleration of (10 m / s 2 , 10 m / s 2 ) from 26 to 50 s, moves at a constant speed from 51 to 75 s, and decelerates with (-10 m / s 2 , -10 m / s 2 ) from 76 to 100 s. In the experiment for verifying the CT grid model, the maneuvering process of the target is set as follows: it makes a turning motion with a turning rate of 4.5*π / 180 (rad / s) from 0 to 20 s, a turning rate of 9*π / 180 (rad / s) from 21 to 50 s, a turning rate of 4.5*π / 180 (rad / s) from 51 to 70 s, a turning rate of 0 from 71 to 80 s, and a turning rate of -4.5*π / 180 (rad / s) from 81 to 100 s.

[0152] Under the above simulation scenarios, the IMM algorithm, the EMA algorithm, and the algorithm proposed by the present invention are compared.

[0153] Refer to Attachment Figure 4 , Figure 5 , Figure 7 and Figure 8 , where Figure 4 and Figure 5 are the trajectory diagrams of the extended target and the overall effect diagrams of the four algorithms in the simulation experiment for the CA grid model; Figure 7 and Figure 8 are the trajectory diagrams of the extended target and the overall effect diagrams of the four algorithms in the simulation experiment for the CT grid model.

[0154] Figure 4 and Figure 7 The abscissa and ordinate of

[0155] Figure 5 and Figure 8 represent the X-axis and Y-axis of the Cartesian coordinate system respectively. The excellent performance of the present invention is demonstrated by comparing the differences between the shape estimation results of the four algorithms for the target in different time periods and the true extended shape.

[0156] Figure 5 and Figure 8The Hausdorff Distance and Root Mean Square Error (RMSE) of the motion state of four algorithms are respectively shown. Among them Figure 5 (a) shows the comparison chart of the Hausdorff Distance of four algorithms in the CA grid simulation experiment, Figure 5 (b), (c), and (d) respectively show the Root Mean Square Error charts of the position, velocity, and acceleration of four algorithms in the CA grid simulation experiment. Figure 8 (a) shows the comparison chart of the Hausdorff Distance of four algorithms in the CT grid simulation experiment, Figure 8 (b) and (c) respectively show the Root Mean Square Error charts of the position and velocity of four algorithms in the CT grid simulation experiment. The Hausdorff Distance and Root Mean Square Error are respectively used to evaluate the shape of the extended target and the estimation accuracy of the motion state. The smaller their values are, the better the tracking accuracy of the algorithm. It can be seen that the algorithm (GSAG-EMA) of the present invention has higher estimation accuracy compared with the other three types of algorithms. From the comprehensive comparison results, whether it is the Hausdorff Distance or the Root Mean Square Error, the estimation performance of the IMM algorithm for such maneuvering extended targets is worse than that of the EMA algorithm, and the shape estimation and motion state estimation performance of the GSAG-EMA algorithm proposed by the present invention are better than those of the EMA algorithm. This is due to the fact that the GSAG-EMA method jointly adjusts the parameters and structure of the grid according to the golden ratio, and at the same time adopts a model transition probability matrix that changes dynamically with time and a hierarchical strategy executed according to the maneuvering level of the object.

[0157] In summary, the present invention uses grid points to represent each model in the model set. The grid points are divided into internal and external grid models, and the parameters of these grid models are adaptively adjusted based on real-time information. The algorithm proposed by the present invention combines the overall and local weighted fusion values of the grid with two models of the internal grid and arranges them spatially according to the golden ratio. By evaluating the difference between the local weighted value and the overall weighted value, the structure and parameters of the grid model are further refined. In order to improve the adaptability in complex maneuvering scenarios, the algorithm updates the transition probability between models in real time and also adopts different algorithm execution strategies according to the maneuvering degree of the target, thereby improving the calculation efficiency. The simulation results show that the algorithm proposed by the present invention significantly improves the estimation accuracy of the motion state and extended shape of the maneuvering extended target.

[0158] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A maneuverable extended target tracking method based on golden section adaptive grid, characterized in that: The steps include: S1. The model set is presented in the form of a grid, which is divided into an internal grid and an external grid. The model is represented as a point in the grid, and the parameters of each model are initialized. S2, using the model probability corresponding to each model in the model set, weighting the model parameters to generate an overall weighted value, i.e., the expected model; S3, calculating the weighted value of the internal grid model, generating local weighting, arranging the overall weighting and the local weighting as well as the two models whose initial positions in the internal grid are located on the left and right sides of the center point of the space according to the ratio of the golden section, and adaptively adjusting the parameters of the models in the internal grid and the external grid, thereby generating a valid model set; S4, according to the model set obtained in step S3, perform IMM estimation on each model, and update the probability and transition probability matrix of each model; S5. According to the size of the model filter residual corresponding to the maximum model probability, the maneuverability of the target is divided into weak maneuverability, medium maneuverability and strong maneuverability, and different algorithm execution strategies are adopted at different maneuverability levels; S6. Use the adjusted model parameters and the updated transfer probability matrix as inputs of the model condition filter and repeat S2-S5.

2. The method for tracking maneuverable extended targets based on golden section adaptive grid according to claim 1, characterized in that: In step S1, the CA model and the CT model are selected as the research objects, and the grid models of the CA model and the CT model are divided into regions and the model parameters are initialized. The parameters of the CA grid model are initialized as follows: Model space A of CA model c is a two-dimensional space, and any point in the space corresponds to a mode: The CA model set contains 9 models. The initial value of the external mesh model is set to: Set the initial values ​​of the inner mesh model to: Among them, A c Represents the model space of the CA model and is a two-dimensional space, a max Indicates the maximum acceleration value that the target can achieve; represents the external mesh model, and Respectively represent the upper edge, lower edge, left edge and right edge models of the initial external mesh model; represents the internal mesh model, and Respectively represent the initial position in the center model of the internal grid Internal mesh models above, below, left and right of ; The parameters of the CT mesh model are initialized as follows: IN c ={w:-w max ≤w≤w max } Where W c represents the model space of the CT model and is a one-dimensional space, w max Indicates the maximum angular velocity that the target can achieve; represents the internal mesh model, and Respectively in the central model Internal mesh models on the left and right; represents the external mesh model, and Represent the left edge and right edge models of the initial external mesh model respectively.

3. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 2, characterized in that: In step S2, for the CA model or the CT model, an expected model is augmented in the original model set. The expected model is a model obtained by performing probability weighted summation of a set of j basic models, and its generation rule is: Among them, M k-1 represents the effective model set at time k-1, is the jth basic model The corresponding probability.

4. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 3, characterized in that: In step S3, the internal grid model is first weighted according to the model probability to obtain the local weighting of the model set; then the overall weighting and the local weighting as well as the two models whose initial positions of the internal grid are located on the left and right sides of the space center point are spatially arranged according to the golden ratio, and the difference in the degree of matching of the two weighted models to the real motion pattern is reflected according to the distance between the two weights, so as to analyze the maneuverability of the target, and adaptively adjust the parameters of the models in the internal grid and the external grid to generate a valid model set.

5. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 4, characterized in that: For the CA model, the parameter adaptation rule of the internal grid is: distance=1.618*dist Rot θ (a,b)=(acosθ-bsinθ,asinθ+bcosθ) in, is the central model of the internal grid and also the local weight of the model set; dist represents the Euclidean distance between the local weight and the overall weight. represents the internal grid model at time k, where and about is centrally symmetrical, so lie in arrive The golden section point of the line segment between lie in arrive At the golden section point between arrive The distance between them is defined as distance; and They are arrive The vector formed is in the positive direction of the x-axis and is located The upper left, upper right, lower left and lower right models of and location; For the CA model, the parameter adaptation rule of the external grid model is: Distance means The distance to the origin, ψ represents The angle between the origin and the positive direction of the x-axis; S1, S2, S3, S4 and S5 represent the five regions after the model space of the CA model is divided; p n (n=1,2,…,5) means The corresponding proportional coefficients when in different areas; and They respectively represent the upper edge, lower edge, left edge and right edge models of the peripheral grid model at time k.

6. The method for tracking maneuverable extended targets based on golden section adaptive grid according to claim 4, characterized in that: The parameter adaptation rule of the internal mesh of the CT model is: in, is the central model of the internal grid and also the local weight of the model set. and about The center is symmetrical, and In arrive At the golden section point of the connecting line, In arrive At the golden section point of the connecting line; and is the internal mesh model of the CT model at time k; The parameter adaptation rule of the external mesh model of the CT model is: Distance1 represents the distance from the expected model of the CT model to the origin; the model space of the CT model is divided into three segments, namely L1, L2 and L3; and Represents the outer mesh model at time k.

7. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 5 or 6, characterized in that: In step S3, the generation rule of the valid model set is: Among them, M k represents the effective model set at time k, and They represent the expected model, internal grid model and external grid model at time k respectively.

8. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 7, characterized in that: In step S4, IMM filtering is performed on each model according to the valid model set obtained in step S3, and each element of the transition probability matrix is ​​adaptively updated after the model probability is updated: in, is the probability of model j after updating at time k, is the correction factor of the element in the i-th row and j-th column of the transfer probability matrix at time k, λ is the correction rate, is the updated transition probability matrix at time k, and r represents the number of models in the model set.

9. The method for tracking a maneuverable extended target based on a golden section adaptive grid according to claim 8, characterized in that: In step S5, the maneuverability of the target is divided into weak maneuverability, medium maneuverability and strong maneuverability according to the size of the model filter residual corresponding to the maximum model probability, and different algorithm execution strategies are adopted at different maneuverability levels, specifically: when the maneuverability level of the target is low, only IMM estimation needs to be performed, and there is no need to update the model parameters and the transfer probability matrix; when the maneuverability of the target is at a medium level, GSAG-EMA is executed while executing IMM estimation, and the transfer probability matrix is ​​not updated; when the maneuverability level of the target is high, IMM estimation, GSAG-EMA and the update of the transfer probability matrix need to be performed simultaneously.

10. The method for tracking maneuverable extended targets based on golden section adaptive grid according to claim 9, characterized in that: In step S5, for model j, its filtering residual is: in, is the new information of model j, is the covariance of the new information, T represents the transposed sign; Assume that the maximum model probability value at time k is: Assume that the filter residual of the model corresponding to the maximum model probability is D k , at different maneuver levels, the execution strategy is: D1 and D2 are two thresholds used to divide the target maneuver level according to the size of the filter residual value.

Images