Target track prediction method and system in dense clutter environment

Abstract

The invention relates to the technical field of multi-target tracking, and particularly discloses a target track prediction method and system in a dense clutter environment, and the method comprises the steps: S1, obtaining an initial measurement value of a sensor and a preset parameter used for limiting the size of a hypothesis tree, and generating the hypothesis tree according to the initial measurement value, the preset parameter and a multi-hypothesis tracking algorithm; assuming that each leaf node of the tree corresponds to a candidate track; s2, each leaf node is preprocessed, and non-repeated candidate tracks are obtained; s3, the candidate tracks are divided into correct tracks and wrong tracks through a neural network classifier, and the number of targets and initial tracks are determined according to the correct tracks; s4, tracking the target according to the initial track, and performing adaptive dynamic adjustment on the number of hypotheses generated in each step in the target tracking process through a multi-hypothesis tracking algorithm of adaptive parameters to obtain hypothesis data associated with the target; and S5, estimating motion parameters of the target through a Kalman filter.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-target tracking, and in particular to a target track prediction method and system in a dense clutter environment. Background Art

[0002] The application of multi-target tracking is very extensive. For example, in the operation of spacecraft and space debris, the estimation of the target's motion state is very important. The multi-target tracking technology can be used to achieve track deduction. In the aerospace field, the predicted track of the target can be used to continuously track spacecraft and space debris, thereby achieving accurate identification; the target track prediction can also achieve real-time scene monitoring, provide effective intelligence and information, etc. Although the application of target track prediction methods based on optical images in fields such as autonomous driving has been relatively mature, the application effect of the method varies greatly under different data sources.

[0003] In particular, traditional methods are limited in the following two scenarios. First, when using radar or infrared sensors for observation, the pixels of the measured images are relatively small and are generally grayscale images, so the image resolution and available information dimension are restricted. Second, when the measurement environment is complex, especially when there is electromagnetic interference and the target's own environment is complex and changeable, the environmental signal-to-noise ratio of multi-target tracking is low, and the influence and interference of many interference factors on the target's own motion are also quite obvious, making it difficult to reuse traditional target track prediction methods.

[0004] Therefore, for specific scenarios with small pixels, grayscale images, and low signal-to-noise ratio, it is difficult to determine which trajectory a certain measurement value comes from. Once the error between the measurement value used and the actual target state is too large, it is easy to cause the target to be lost, which in turn leads to target tracking and recognition failure.

[0005] In fact, the existing target track prediction methods can achieve good results in scenarios with high signal-to-noise ratio and sparse targets. However, with the widespread application of various modern sensors and various high-precision technologies, as well as the development of stealth technology and jamming technology, the measurement data obtained has increased dramatically. At the same time, the interference, clutter and noise in the actual environment have become more complex and changeable. Dense clutter mixed in the measurement data leads to a decrease in the accuracy of target tracking, and may cause problems such as "combination explosion" and target loss, which in turn affects the subsequent tracking and identification decisions. Summary of the invention

[0006] In view of the above problems, the purpose of the present invention is to provide a target track prediction method and system in a dense clutter environment, which integrates a multi-hypothesis tracking algorithm with the track initiation to reduce the false alarm rate and missed detection rate in the track initiation stage; and realizes adaptive adjustment of the parameters of the multi-hypothesis tracking algorithm, which effectively reduces the calculation time while ensuring accuracy.

[0007] The present invention provides a target track prediction method in a dense clutter environment, comprising:

[0008] S1, obtaining initial measurement values ​​of the sensor and preset parameters for limiting the size of the hypothesis tree, and generating a hypothesis tree according to the initial measurement values, the preset parameters and a multi-hypothesis tracking algorithm; each leaf node of the hypothesis tree corresponds to a candidate track;

[0009] S2, preprocessing each leaf node to obtain non-repetitive candidate tracks;

[0010] S3, classifying the candidate tracks into correct tracks and incorrect tracks by a neural network classifier, and determining the number of targets and the initial track according to the correct tracks;

[0011] S4, tracking the target according to the initial track, adaptively and dynamically adjusting the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters, and obtaining hypothesis data associated with the target;

[0012] S5, estimating the motion parameters of the target through the Kalman filter;

[0013] S6, obtaining a target predicted track according to the motion parameters, the initial track and the assumed data.

[0014] In a possible implementation, the S1 includes:

[0015] Based on the N-scan pruning method and / or the M-optimal MHT method, a hypothesis tree is generated according to the initial measurement value, the preset parameters and a multi-hypothesis tracking algorithm.

[0016] In a possible implementation, the neural network classifier is a five-layer neural network, using a RELU function as an activation function; S3 includes:

[0017] Extracting the velocity feature, acceleration feature and angular velocity feature of each candidate track;

[0018] Inputting the velocity feature, acceleration feature and angular velocity feature into the neural network classifier to obtain the score of each candidate track in each category output by the neural network classifier;

[0019] The category with the highest score is used as the final category of the candidate track; the categories include: correct track and wrong track.

[0020] In a possible implementation, S3 further includes:

[0021] Determining the speed characteristic according to the average speed between two consecutive time points of the candidate track;

[0022] Determine the acceleration feature according to the acceleration between two connecting lines formed by three consecutive time points of the candidate track;

[0023] The angular velocity characteristic is determined according to the angle between two connecting lines formed by three consecutive time points of the candidate track.

[0024] In a possible implementation, the S4 includes:

[0025] Sort the multiple hypotheses generated at each moment of the target by confidence, and take the hypothesis with the highest confidence as the optimal hypothesis;

[0026] When the last hypothesis in the sorting is not used to update the target within a preset time period, the last hypothesis in the sorting is deleted.

[0027] In a possible implementation, the S5 includes:

[0028] The motion parameters of the target are estimated according to the following formula:

[0029] x(k+1|k)=F(k)x(k|k)+G(k)u(k)

[0030] z(k+1|k)=H(k+1)x(k+1|k)

[0031] P(k+1|k)=F(k)P(k|k)F(k) T +Q(k)

[0032] Among them, P(k+1|k) is the predicted covariance of the state at time k+1, F(k) is the state transfer matrix, P(k|k) is the estimated covariance of the state at time k, Q(k) is the covariance of the process noise, z(k+1|k) is the predicted value of the measurement at time k+1, H(k+1) is the measurement matrix of the target at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, x(k|k) is the estimated value of the target state at time k, G(k) is the input matrix, and u(k) is the control quantity.

[0033] In a possible implementation, S5 further includes:

[0034] Update the target's motion parameters according to the following formula:

[0035] S(k+1|k)=H(k+1)P(k+1|k)H(k+1) T +R(k+1)

[0036] K(k+1)=P(k+1|k)H(k+1) T S(k+1) -1

[0037] x(k+1|k+1)=x(k+1|k)+K(k+1)[z(k+1)-z(k+1|k)]

[0038] P(k+1|k+1)=P(k+1|k)-K(k+1)S(k+1)K(k+1) T

[0039] Among them, S(k+1|k) is the predicted covariance of the observation at time k+1, H(k+1) is the measurement matrix of the target at time k+1, P(k+1|k) is the predicted covariance of the state at time k+1, R(k+1) is the covariance of the measurement noise at time k+1, K(k+1) is the Kalman filter gain at time k+1, x(k+1|k+1) is the estimated value of the target state at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, z(k+1) is the measured value of the target at time k+1, z(k+1|k) is the predicted value of the measurement of the target at time k+1, and P(k+1|k+1) is the estimated covariance of the state at time k+1.

[0040] In one possible implementation, the state equation of the target is as follows:

[0041] x(k+1)=F(k)x(k)+G(k)u(k)+v(k)

[0042] Among them, x(k+1) is the state vector of the target at time k+1, x(k) represents the state vector of the target at time k, F(k) is the state transfer matrix, G(k) is the input matrix, v(k) represents the process noise sequence, and u(k) is the control quantity.

[0043] In a possible implementation, the measurement equation of the target is as follows:

[0044] z(k)=H(k)x(k)+w(k)

[0045] Where z(k) is the measurement vector of the target at time k, H(k) is the measurement matrix of the target at time k, x(k) is the state vector of the target at time k, and w(k) is the measurement noise sequence.

[0046] The present invention also provides a target track prediction system in a dense clutter environment, which is used to implement any of the above target track prediction methods, including:

[0047] A hypothesis tree generation module is used to obtain initial measurement values ​​of the sensor and preset parameters for limiting the size of the hypothesis tree, and generate a hypothesis tree according to the initial measurement values, the preset parameters and a multi-hypothesis tracking algorithm; each leaf node of the hypothesis tree corresponds to a candidate track;

[0048] A candidate track generation module is used to pre-process each leaf node to obtain non-repetitive candidate tracks;

[0049] A neural network classifier module, used for classifying the candidate tracks into correct tracks and incorrect tracks by a neural network classifier, and determining the number of targets and the initial track according to the correct tracks;

[0050] An adaptive adjustment module is used to track the target according to the initial track, and adaptively and dynamically adjust the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters to obtain hypothesis data associated with the target;

[0051] A Kalman filter module, used for estimating the motion parameters of the target through a Kalman filter;

[0052] A prediction module is used to obtain a target predicted track according to the motion parameters, the initial track and the hypothetical data.

[0053] The target track prediction method and system provided by the present invention in a dense clutter environment, in the process of multi-target tracking, uses all the measurement value information of the previous few moments to establish a hypothesis tree of the multi-hypothesis tracking algorithm, and each branch of the hypothesis tree forms a candidate track, which is output to the track initiation neural network classifier to obtain the number of targets and the initial track of the target. The initial track is used as the initial value of the target motion parameter, and a multi-hypothesis tracking algorithm with adaptive parameters is used to perform data association and state estimation to obtain the predicted track of the target. On the one hand, the multi-hypothesis tracking algorithm is integrated with the track initiation to reduce the false alarm rate and missed detection rate in the track initiation stage; on the other hand, the adaptive adjustment of the parameters of the multi-hypothesis tracking algorithm is realized, which effectively reduces the operation time while ensuring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic diagram of a target track prediction method provided by an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of a hypothesis tree provided for an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of the structure of a neural network classifier provided in an embodiment of the present invention;

[0057] Figure 4 A schematic diagram of a multi-hypothesis tracking algorithm with adaptive parameters provided by an embodiment of the present invention;

[0058] Figure 5 A comparison chart of the track initial missed detection rate provided by an embodiment of the present invention and the logic method;

[0059] Figure 6A comparison chart of the track start false alarm rate provided by an embodiment of the present invention and the logic method;

[0060] Figure 7 The root mean square error between the predicted tracks and the actual tracks of the 15 targets provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.

[0062] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0063] Figure 1 A flow chart of a target track prediction method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a target track prediction method in a dense clutter environment, comprising:

[0064] S1, obtaining the initial measurement values ​​of the sensor and the preset parameters for limiting the size of the hypothesis tree, and generating a hypothesis tree according to the initial measurement values, the preset parameters and the multi-hypothesis tracking algorithm; each leaf node of the hypothesis tree corresponds to a candidate track;

[0065] In a possible implementation, S1 includes: generating a hypothesis tree according to initial measurement values, preset parameters and a multi-hypothesis tracking algorithm based on an N-scan pruning method and / or an M-optimal MHT method.

[0066] The multi-hypothesis tracking algorithm believes that subsequent data will solve the uncertainty of previous data association. Its principle is to postpone the decision of uncertain data association problems. That is, when there are multiple possible measurement values ​​in the target's wave gate, several candidate hypotheses are established first, and then the measurement values ​​associated with the previous steps are determined after more data and information are obtained for each subsequent target.

[0067] The generation of hypothesis is the core part of the multi-hypothesis tracking algorithm. Let Ω(k) = {Ω i (k),i=1,2,...,I k} represents the hypothesis set containing all hypotheses at time k, Z(k) is the set of measured values ​​in the target gate at time k, and Z krepresents the cumulative set of confirmed measurements in the wave gate from the initial time to time k. The assumption in Ω(k) is to convert Z k The hypothesis at time k is obtained by associating Ω(k-1) at time k-1 with the measurement set Z(k) at time k. In hypothesis generation, the amount of computation and storage required by the MHT method grows exponentially. Therefore, in practical applications, it is generally necessary to reduce the amount of computation through pruning methods, including N-scan pruning method, M-optimal MHT method, etc.

[0068] The main idea of ​​the N-scan method is to reduce the number of hypotheses by controlling the depth of the hypothesis tree, that is, the uncertainty of data association at time k will be resolved at time k+N. The main steps of the N-scan pruning method are to reduce the number of hypotheses by controlling the depth of the hypothesis tree.

[0069]

[0070] in is the posterior probability of the parent node hypothesis, P D is the probability of the target being detected, β C is the clutter density, β N is the density of the new target, N DT is the number of measurements associated with an existing target, N T is the number of existing target trajectories, N C is the number of measurements judged as clutter, N N is the number of new trajectories. The probability of a measurement value being associated with a target satisfies the normal distribution N, is the predicted measurement value of the jth target at time k, S j is the covariance.

[0071] The M-optimal MHT method mainly reduces the number of hypotheses by controlling the width of the hypothesis tree. At each step, M optimal hypotheses are generated as leaf nodes of the hypothesis tree. Generating M optimal hypotheses can be transformed into a bipartite graph matching problem. According to the cost matrix, the M allocations with the lowest cost, i.e., the M optimal association hypotheses, are obtained, and these hypotheses are sorted by probability.

[0072] If the N-scan method and the M-optimal method are used at the same time, assuming that the maximum width of the tree is controlled at M N-1 , which can greatly reduce storage requirements and computational workload.

[0073] The initial measurement value is input into the multi-hypothesis tracking algorithm. At this time, the number of targets and the initial values ​​of target motion parameters are unknown. Each measurement value is regarded as a possible target, and a hypothesis tree of a certain depth and width is generated according to the set parameters. Unlike the complete multi-hypothesis tracking algorithm, after the hypothesis tree is established, no optimal hypothesis decision is made. At this time, each leaf node of the hypothesis tree corresponds to a candidate track. Figure 2 A schematic diagram of a hypothesis tree provided for an embodiment of the present invention, i.e., a hypothesis tree at time K, wherein parameters M and N can be adjusted according to the actual memory and computing power used. The larger the parameters M and N, the better the effect of the track start, but the memory and computing resources consumed will also be greater.

[0074] S2, pre-process each leaf node to obtain non-repetitive candidate tracks;

[0075] Assuming that the number of candidate tracks is R, the candidate tracks can be expressed as: Z = {r1, r2, …, r K}

[0076] Among them, r i is the position vector at time i, expressed in rectangular coordinates, and K means that the work of starting the track is completed within time K.

[0077] S3, classifying the candidate tracks into correct tracks and incorrect tracks through a neural network classifier, and determining the number of targets and the initial track according to the correct tracks;

[0078] Figure 3 The schematic diagram of the structure of the neural network classifier provided by the embodiment of the present invention is a five-layer neural network, which uses the RELU function as the activation function, and uses a fully connected network internally. The number of neurons in each layer and the number of hidden layers are network hyperparameters, which can be changed according to the actual situation and are determined by methods such as cross-validation in actual use. The size of the input matrix is ​​R×(3×K-5), which represents the feature vectors of different tracks, and the output size is R×2, which represents the score of each track on the correct track and the wrong track. The final classification result of each track is the one with the highest score.

[0079] In a possible implementation, S3 includes: extracting speed features, acceleration features, and angular velocity features of each candidate track; inputting the speed features, acceleration features, and angular velocity features into a neural network classifier to obtain scores of each candidate track in each category output by the neural network classifier; taking the category with the highest score as the final category of the candidate track; the categories include: correct tracks and incorrect tracks.

[0080] In a possible implementation, S3 further includes:

[0081] Determine the speed feature v based on the average speed between two consecutive time points of the candidate track i ;

[0082] Determine the acceleration feature a based on the acceleration between two lines formed by three consecutive time points of the candidate track i ;

[0083] Determine the angular velocity feature θ based on the angle between two lines formed by three consecutive time points of the candidate track i .

[0084]

[0085] The above features can be expressed as x r =(v1,…,v K-1 ,a1,…,a K-2 ,θ1,…,θ K-2 ),r=1,2,…,R.

[0086] S4, tracking the target according to the initial track, adaptively and dynamically adjusting the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters, and obtaining hypothesis data associated with the target;

[0087] In a possible implementation, S4 includes: sorting multiple hypotheses generated for the target at each moment according to confidence, and taking the hypothesis with the highest confidence as the optimal hypothesis; when the last hypothesis in the sorting is not used to update the target within a preset time, deleting the last hypothesis in the sorting.

[0088] Figure 4 A schematic diagram of a multi-hypothesis tracking algorithm with adaptive parameters provided by an embodiment of the present invention, the main idea of ​​the adaptive multi-hypothesis tracking algorithm is to adaptively and dynamically adjust the number of hypotheses M generated at each step during the tracking process. The specific implementation method is to use a sliding window method to determine whether a target can be stably tracked using a smaller M. The hypotheses generated at each moment are arranged from large to small according to their confidence, that is, the one at the front is the best one. If the Mth hypothesis is not used to update the target within a continuous period of time l, it means that reducing M at this time can achieve better results, and the reduced M is used for target tracking later.

[0089] S5, estimating the motion parameters of the target through the Kalman filter;

[0090] In a possible implementation, S5 includes:

[0091] The motion parameters of the target are estimated according to the following formula:

[0092] x(k+1|k)=F(k)x(k|k)+G(k)u(k)

[0093] z(k+1|k)=H(k+1)x(k+1|k)

[0094] P(k+1|k)=F(k)P(k|k)F(k) T +Q(k)

[0095] Among them, P(k+1|k) is the predicted covariance of the state at time k+1, F(k) is the state transfer matrix, P(k|k) is the estimated covariance of the state at time k, Q(k) is the covariance of the process noise, z(k+1|k) is the predicted value of the measurement at time k+1, H(k+1) is the measurement matrix of the target at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, x(k|k) is the estimated value of the target state at time k, G(k) is the input matrix, and u(k) is the control quantity.

[0096] In a possible implementation, S5 further includes:

[0097] Update the target's motion parameters according to the following formula:

[0098] S(k+1|k)=H(k+1)P(k+1|k)H(k+1) T +R(k+1)

[0099] K(k+1)=P(k+1|k)H(k+1) T S(k+1) -1

[0100] x(k+1|k+1)=x(k+1|k)+K(k+1)[z(k+1)-z(k+1|k)]

[0101] P(k+1|k+1)=P(k+1|k)-K(k+1)S(k+1)K(k+1) T

[0102] Among them, S(k+1|k) is the predicted covariance of the observation at time k+1, H(k+1) is the measurement matrix of the target at time k+1, P(k+1|k) is the predicted covariance of the state at time k+1, R(k+1) is the covariance of the measurement noise at time k+1, K(k+1) is the Kalman filter gain at time k+1, x(k+1|k+1) is the estimated value of the target state at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, z(k+1) is the measured value of the target at time k+1, z(k+1|k) is the predicted value of the measurement of the target at time k+1, and P(k+1|k+1) is the estimated covariance of the state at time k+1.

[0103] In one possible implementation, the state equation of the target is as follows:

[0104] x(k+1)=F(k)x(k)+G(k)u(k)+v(k)

[0105] Among them, x(k+1) is the state vector of the target at time k+1, x(k) represents the state vector of the target at time k, F(k) is the state transfer matrix, G(k) is the input matrix, v(k) represents the process noise sequence, which is a Gaussian white noise with a mean equal to zero, and u(k) is the control variable.

[0106] In a possible implementation, the measurement equation of the target is as follows:

[0107] z(k)=H(k)x(k)+w(k)

[0108] Where z(k) is the measurement vector of the target at time k, H(k) is the measurement matrix of the target at time k, x(k) is the state vector of the target at time k, and w(k) is the measurement noise sequence.

[0109] S6, obtaining the target predicted track according to the motion parameters, the initial track and the assumed data.

[0110] The present invention also provides a target track prediction system in a dense clutter environment, which is used to implement any of the above target track prediction methods, including:

[0111] A hypothesis tree generation module is used to obtain the initial measurement values ​​of the sensor and the preset parameters for limiting the size of the hypothesis tree, and generate a hypothesis tree according to the initial measurement values, the preset parameters and the multi-hypothesis tracking algorithm; each leaf node of the hypothesis tree corresponds to a candidate track;

[0112] The candidate track generation module is used to pre-process each leaf node to obtain non-repetitive candidate tracks;

[0113] A neural network classifier module is used to classify candidate tracks into correct tracks and incorrect tracks through a neural network classifier, and determine the number of targets and the initial track according to the correct track;

[0114] An adaptive adjustment module is used to track the target according to the initial track, and to adaptively and dynamically adjust the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters to obtain hypothesis data associated with the target;

[0115] A Kalman filter module, used for estimating the motion parameters of the target through a Kalman filter;

[0116] The prediction module is used to obtain the target predicted track based on the motion parameters, the initial track and the assumed data.

[0117] Specifically, the target track prediction method of the present invention is used to simulate the track prediction of 15 targets in environments with different clutter densities. The simulation results are as follows: Figure 5-7 express. Figure 5 A comparison chart of the track initial missed detection rate provided by an embodiment of the present invention and the logic method, Figure 6 A comparison chart of the track start false alarm rate provided by an embodiment of the present invention and the logic method, Figure 7 The root mean square error between the predicted tracks and the actual tracks of the 15 targets provided by the embodiment of the present invention. The simulation results show that the present invention combines the multi-hypothesis tracking algorithm with the machine learning track initiation, reduces the amount of calculation of the traditional track initiation method, and reduces the missed detection rate and false alarm rate; in addition, it realizes the adaptive adjustment of the parameters of the multi-hypothesis tracking algorithm, effectively shortens the operation time while ensuring the correlation accuracy, and realizes efficient and accurate target track prediction.

[0118] The target track prediction method and system provided by the present invention in a dense clutter environment, in the process of multi-target tracking, uses all the measurement value information of the previous few moments to establish a hypothesis tree of the multi-hypothesis tracking algorithm, and each branch of the hypothesis tree forms a candidate track, which is output to the track initiation neural network classifier to obtain the number of targets and the initial track of the target. The initial track is used as the initial value of the target motion parameter, and a multi-hypothesis tracking algorithm with adaptive parameters is used to perform data association and state estimation to obtain the predicted track of the target. On the one hand, the multi-hypothesis tracking algorithm is integrated with the track initiation to reduce the false alarm rate and missed detection rate in the track initiation stage; on the other hand, the adaptive adjustment of the parameters of the multi-hypothesis tracking algorithm is realized, which effectively reduces the operation time while ensuring accuracy.

[0119] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A target track prediction method in a dense clutter environment, characterized in that: include:: S1, obtaining initial measurement values ​​of a sensor and preset parameters for limiting the size of a hypothesis tree, and generating a hypothesis tree according to the initial measurement values, the preset parameters and a multi-hypothesis tracking algorithm; Each leaf node of the hypothesis tree corresponds to a candidate track; S2, preprocessing each leaf node to obtain non-repetitive candidate tracks; S3, classifying the candidate tracks into correct tracks and incorrect tracks by a neural network classifier, and determining the number of targets and the initial track according to the correct tracks; S4, tracking the target according to the initial track, adaptively and dynamically adjusting the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters, and obtaining hypothesis data associated with the target; S5, estimating the motion parameters of the target through the Kalman filter; S6, obtaining a target predicted track according to the motion parameters, the initial track and the assumed data.

2. The target track prediction method according to claim 1, characterized in that: The S1 includes: Based on the N-scan pruning method and / or the M-optimal MHT method, a hypothesis tree is generated according to the initial measurement value, the preset parameters and a multi-hypothesis tracking algorithm.

3. The target track prediction method according to claim 1, characterized in that: The neural network classifier is a five-layer neural network, using the RELU function as an activation function; S3 includes: Extracting the velocity feature, acceleration feature and angular velocity feature of each candidate track; Inputting the velocity feature, acceleration feature and angular velocity feature into the neural network classifier to obtain the score of each candidate track in each category output by the neural network classifier; The category with the highest score is used as the final category of the candidate track; the categories include: correct track and wrong track.

4. The target track prediction method according to claim 3, characterized in that: The S3 further includes: Determining the speed characteristic according to the average speed between two consecutive time points of the candidate track; Determine the acceleration feature according to the acceleration between two connecting lines formed by three consecutive time points of the candidate track; The angular velocity characteristic is determined according to the angle between two connecting lines formed by three consecutive time points of the candidate track.

5. The target track prediction method according to claim 1, characterized in that: The S4 includes: Sort the multiple hypotheses generated at each moment of the target by confidence, and take the hypothesis with the highest confidence as the optimal hypothesis; When the last hypothesis in the sorting is not used to update the target within a preset time period, the last hypothesis in the sorting is deleted.

6. The target track prediction method according to claim 1, characterized in that: The S5 includes: The motion parameters of the target are estimated according to the following formula: x(k+1|k)=F(k)x(k|k)+G(k)u(k) z(k+1|k)=H(k+1)x(k+1|k) P(k+1|k)=F(k)P(k|k)F(k) T +Q(k) Among them, P(k+1|k) is the predicted covariance of the state at time k+1, F(k) is the state transfer matrix, P(k|k) is the estimated covariance of the state at time k, Q(k) is the covariance of the process noise, z(k+1|k) is the predicted value of the measurement at time k+1, H(k+1) is the measurement matrix of the target at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, x(k|k) is the estimated value of the target state at time k, G(k) is the input matrix, and u(k) is the control quantity.

7. The target track prediction method according to claim 6, characterized in that: The S5 further includes: Update the target's motion parameters according to the following formula: S(k+1|k)=H(k+1)P(k+1|k)H(k+1) T +R(k+1) K(k+1)=P(k+1|k)H(k+1) T S(k+1) -1 x(k+1|k+1)=x(k+1|k)+K(k+1)[z(k+1)-z(k+1|k)] P(k+1|k+1)=P(k+1|k)-K(k+1)S(k+1)K(k+1) T Among them, S(k+1|k) is the predicted covariance of the observation at time k+1, H(k+1) is the measurement matrix of the target at time k+1, P(k+1|k) is the predicted covariance of the state at time k+1, R(k+1) is the covariance of the measurement noise at time k+1, K(k+1) is the Kalman filter gain at time k+1, x(k+1|k+1) is the estimated value of the target state at time k+1, x(k+1|k) is the predicted value of the target state at time k+1, z(k+1) is the measured value of the target at time k+1, z(k+1|k) is the predicted value of the measurement of the target at time k+1, and P(k+1|k+1) is the estimated covariance of the state at time k+1.

8. The target track prediction method according to claim 6, characterized in that: The state equation of the target is as follows: x(k+1)=F(k)x(k)+G(k)u(k)+v(k) Among them, x(k+1) is the state vector of the target at time k+1, x(k) represents the state vector of the target at time k, F(k) is the state transfer matrix, G(k) is the input matrix, v(k) represents the process noise sequence, and u(k) is the control quantity.

9. The target track prediction method according to claim 6, characterized in that: The measurement equation of the target is as follows: z(k)=H(k)x(k)+w(k) Where z(k) is the measurement vector of the target at time k, H(k) is the measurement matrix of the target at time k, x(k) is the state vector of the target at time k, and w(k) is the measurement noise sequence.

10. A target track prediction system in a dense clutter environment, used to implement the target track prediction method according to any one of claims 1 to 9, characterized in that: include: A hypothesis tree generation module, used for obtaining initial measurement values ​​of the sensor and preset parameters for limiting the size of the hypothesis tree, and generating a hypothesis tree according to the initial measurement values, the preset parameters and a multi-hypothesis tracking algorithm; Each leaf node of the hypothesis tree corresponds to a candidate track; A candidate track generation module is used to pre-process each leaf node to obtain non-repetitive candidate tracks; A neural network classifier module, used for classifying the candidate tracks into correct tracks and incorrect tracks by a neural network classifier, and determining the number of targets and the initial track according to the correct tracks; An adaptive adjustment module is used to track the target according to the initial track, and adaptively and dynamically adjust the number of hypotheses generated in each step of the target tracking process through a multi-hypothesis tracking algorithm with adaptive parameters to obtain hypothesis data associated with the target; A Kalman filter module, used for estimating the motion parameters of the target through a Kalman filter; A prediction module is used to obtain a target predicted track according to the motion parameters, the initial track and the hypothetical data.

Images