Multi-target interaction target tracking method based on gated neural network

Abstract

The invention belongs to the technical field of target tracking, and particularly relates to a multi-target interaction target tracking method based on a gated neural network. Comprising the following steps: network design: two gating units containing neural network modules and a Kalman tracking framework are included; wherein the first-level gating unit serves as a memory updating module and is used for acquiring a memory factor dk capable of representing interactive information implied in a multi-target historical state; the second-level gating unit serves as an interactive input module and is used for decoding interactive information in the memory factor dk, converting the interactive information into a state space uk and fusing the state space uk into a Kalman tracking framework; network training: training all learnable parameters in the network through a loss function between an estimation state and a truth value state by using multi-target historical state data; and multi-target interaction tracking: obtaining a target state by using the trained neural network, and realizing multi-target interaction tracking.

Description

Technical Field

[0001] This invention belongs to the field of target tracking technology, specifically relating to a multi-target interactive target tracking method based on a gated neural network. Background Technology

[0002] With the rapid development of smart city transportation technologies, traffic radar target tracking has become a key research area. Therefore, accurate target tracking technology plays a crucial role in improving traffic flow and reducing accident rates. However, in dynamic and dense traffic environments, it is necessary to consider not only the tracking of individual targets but also the interactions between multiple targets, such as lane changes, pedestrian avoidance, vehicle following, intersection merging, and sudden braking. During target interactions, clutter from non-target signals (such as road signs, pedestrians, and other vehicles) can interfere with target signals, reducing the accuracy of radar detection and causing fluctuations in measurement data. This indicates that a specific target may exhibit a consistently high false negative rate from a certain point in time, posing a challenge to the tracking algorithm's ability to acquire target measurement data, thus affecting the continuity and stability of tracking.

[0003] Therefore, comprehensively considering the interaction information between multiple targets is crucial for improving tracking performance. These interactions occur frequently in complex traffic scenarios and significantly affect the state evolution of different targets. These interactions may involve factors such as relative speed, vehicle spacing, and movement intentions between multiple targets, but traditional models struggle to accurately characterize these complex relationships. To construct effective multi-vehicle interaction models, car following (CF) models based on the longitudinal interaction between two vehicles have become one of the mainstream methods. Among them, the Gazis-Herman-Rothery (GHR) model, as a classic method, assumes that the acceleration of the following vehicle is calculated from the relative speed and distance between the preceding and following vehicles, while also considering the driver's reaction delay. This model can effectively capture the dynamic interactions between vehicles and characterize the relationship between acceleration through relative speed, distance, and reaction time. To address the aforementioned problem of measurement data fluctuations, integrating multi-target interaction information can more accurately simulate the motion state of target vehicles, thereby improving tracking accuracy.

[0004] Traditional model-driven filtering methods and purely data-driven neural network (NN) methods each have their limitations in addressing these challenges. First, the performance of traditional Kalman filtering (KF) methods is highly dependent on the accurate modeling of target states and noise statistics. In real-world traffic scenarios, the complex interactions between multiple targets make building accurate and concise models a significant challenge. In contrast, rapidly developing neural network methods, especially recurrent neural networks (RNNs), have significant advantages in processing time-series data. RNNs, through their recurrent architecture, preserve the temporal information of the input sequence, demonstrating a significant advantage in capturing complex temporal dependencies. In particular, variants such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) overcome the gradient vanishing and gradient exploding problems in traditional RNNs through gating mechanisms, significantly improving their performance in long-term time-series data analysis tasks. These NN methods rely on large amounts of data for training and can, to some extent, resolve complex nonlinear relationships. However, these methods typically face problems such as poor interpretability, large data requirements, and high computational complexity.

[0005] To address these challenges, researchers have begun exploring hybrid approaches that combine traditional Kalman filtering (KF) techniques with neural networks (NNs) to improve target tracking efficiency and accuracy by integrating the advantages of both. This hybrid model combines the model-driven nature of KF with the data-driven capabilities of NNs, offering higher interpretability and performance when handling complex nonlinear relationships. Specifically, RNNs (such as LSTM or GRU) can be fused into the KF framework to extract dynamic system features and complex measurement data fluctuation patterns. This fusion allows the filter to maintain high tracking performance even when measurement data fluctuates. Guy Revach et al. proposed KalmanNet based on the Kalman filtering framework, using neural networks to learn the filter gain in the Kalman filter and creatively designing two filter gain calculation networks based on the iterative approach of the covariance matrix in the Kalman filter. Furthermore, in previous work by Yan et al., they designed an explainable gated Bayesian RNN (EGBRNN). The internal structure of its gated network is based on Bayesian filtering (BF) theory to integrate prior model information and offline data, thereby reducing parameter requirements and improving data efficiency. This method focuses on the temporal correlation of single-target state evolution but neglects the interaction process between multiple targets. This interaction occurs frequently in high-density traffic scenarios and mutually influences the state evolution of different targets. By learning the multi-target interaction relationships in traffic radar data, the accuracy of target tracking can be improved when measurement data fluctuates.

[0006] To address the decline in tracking accuracy and stability caused by fluctuations in measurement data in traffic scenarios, this invention proposes a novel GRNN for Multi-Target Interaction (GRNN-MTI) based on multi-target interaction. First, the state prediction model in the KF framework is optimized using acceleration control terms learned by the NN module. Second, latent interaction information in the historical trajectories of multiple targets is mined, and a memory factor is introduced. Simultaneously, a non-Markov model is transformed into an equivalent first-order Markov model with memory through nested functions, enabling the GRNN to decode accumulated interaction information and generate specified interaction acceleration terms. Then, based on the KF framework, a GRNN network structure with a two-layer gating unit at its core is designed. This structure includes two key components: an interaction memory update gate (IMUG) to capture historical patterns during multi-target interactions; and an interaction input gate (IIG) to transform the memorized interaction data into acceleration control terms for the motion model. Even if a target continuously lacks measurement data input after a certain moment, this method can still achieve stable tracking by effectively utilizing multi-target interaction information. Through a carefully designed gating unit structure, this method successfully integrates offline data with prior model knowledge, exhibiting high interpretability. Finally, experimental results show that, even when measurement data for a specific target is consistently lacking, this method significantly outperforms traditional KF and advanced NN filtering methods in tracking performance, effectively ensuring the continuity and stability of the tracking process. Summary of the Invention

[0007] This invention proposes a multi-target interactive tracking method based on gated neural networks. This method effectively improves tracking performance in complex traffic scenarios by integrating the interpretable computational architecture of KF with the memory iteration and nonlinear fitting capabilities of GRNN.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention provides a multi-target interactive tracking method based on a gated neural network, comprising the following steps:

[0010] Network design: includes two gating units containing neural network modules and a Kalman tracking framework;

[0011] The first-level gating unit, acting as a memory update module, is used to acquire the memory factor d, which represents the interaction information implicit in the historical states of multiple targets. k The second-level gating unit serves as an interactive input module, used for controlling the memory factor d. k The interactive information in the data is decoded and converted into a state space u.k Integrating the Kalman tracking framework;

[0012] The Kalman tracking framework includes a Kalman prediction model and a Kalman update model. The Kalman prediction model is used to determine the state space u. k The target state and covariance matrix are predicted; the Kalman update model updates the predicted target state and covariance matrix when it receives measurement data.

[0013] Network training: Using multi-objective historical state data, all learnable parameters in the network are trained by estimating the loss function between the current state and the true state;

[0014] Multi-target interactive tracking: Using a trained neural network to obtain the target state, multi-target interactive tracking can be achieved.

[0015] Optionally, the first-level neural network of the present invention serves as a memory update module: based on the interaction target at time k-1 in the filtered state. Difference vector and memory factor d k-1 Predict the memory factor d at time k k .

[0016] Optionally, the memory factor d described in this invention k for:

[0017] d k =W d2 (σ(W d1 i d +b d1 ))+b d2

[0018]

[0019] Among them, W d and b d d respectively k The weights and biases, W d =(W d1 W d2 ) and b d =(b d1 ,b d2 ) represents the set of learnable parameters, σ() is the hyperbolic tangent activation function, con{} represents the concatenation function, and Ψ(·) is the maximum normalization function.

[0020] Optionally, the second-level neural network of the present invention serves as an interactive input module: based on the filtered states of the other N-1 interactive targets (excluding the i-th target) at time k-1. and memory factor d kPredict the state space u at time k k .

[0021] Optionally, the present invention described

[0022]

[0023] Among them, W u and b u for The weights and biases, W u =(W u1 W u2 ) and b u =(b u1 ,b u2 ) represents the set of learnable parameters, tanh() is the hyperbolic tangent activation function, con{} represents the concatenation function, and Ψ(·) is the maximum normalization function.

[0024] Optionally, the Kalman prediction model of this invention is:

[0025]

[0026] P k|k-1 =F k P k-1|k-1 (F k ) T +Q k

[0027] in, This represents the predicted value of the target state at time k. F represents the predicted value of the target state at time k-1. k Let G represent the state transition matrix. k P represents the control matrix. k|k-1 Let P represent the predicted covariance matrix at time k. k-1|k-1 Let Q represent the predicted covariance matrix at time k-1. k The covariance matrix represents the process noise.

[0028] Optionally, the Kalman update model described in this invention is:

[0029]

[0030] P k|k =(IK k H k )P k|k-1

[0031] Among them, H k Measurement matrix, R k The covariance matrix of the measurement noise, z kK represents the measurement data. k This represents the Kalman gain.

[0032] Optionally, the present invention uses the minimum mean square error (MSE) loss function constructed with L2 regularization to train the network based on the minimum mean square error criterion of state estimation. The MSE loss function is defined as follows:

[0033]

[0034] Where, ω decay This is the scaling factor for adjusting the regularization term, where i = {1, 2, ..., N} are the indices of the N interaction targets, m = {1, 2, ..., M} are the indices of the M training samples, and k = {1, 2, ..., T} are the values ​​of the regularization term. len / T} is a length of T len (s) is the index of the time-sampling data, where Ω is the set of learnable parameters.

[0035] Optionally, this invention employs mini-batch stochastic gradient descent to train and update Ω.

[0036] Secondly, the present invention provides a multi-target interactive tracking device based on a gated neural network, comprising:

[0037] The memory update module is used to obtain the memory factor d, which can represent the interaction information implicit in the historical states of multiple objectives. k ;

[0038] An interactive input module is used for the memory factor d. k The interactive information in the data is decoded and converted into a state space u. k ;

[0039] Kalman prediction model, used to predict based on the state space u k Predict the target state and covariance matrix;

[0040] The Kalman update model updates the predicted target state and covariance matrix when measurement data is received.

[0041] Beneficial effects:

[0042] First, this invention specifically designs a multi-target interactive target tracking method based on a gated neural network for traffic radar target tracking when measurement data fluctuates. This provides richer multi-target interactive information for target tracking in complex traffic scenarios, enabling the algorithm to estimate the target state more accurately.

[0043] Secondly, this invention transforms and integrates multi-target interaction information into the KF framework, which has good interpretability, low complexity and high tracking accuracy. Each gate unit in the designed multi-target interactive target tracking method based on gated neural network has a clear and targeted function. By integrating the prior model of KF with the multi-target interaction information in offline data, the decrease in tracking accuracy when measurement data fluctuates is effectively mitigated.

[0044] Third, the present invention uses experimental data to train and verify the designed network. The results show that the proposed algorithm is superior to the traditional KF method and the current advanced data-driven NN filtering method, and can achieve high estimation accuracy while reducing the number of learnable parameters. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments 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.

[0046] Figure 1 The network structure diagram designed for the method of this invention;

[0047] Figure 2 This is a coordinate diagram of the tracking position of the two interactive targets in the method of the present invention;

[0048] Figure 3 This is a coordinate diagram of the tracking positions of the three interactive targets in the method of this invention;

[0049] Figure 4 This is a coordinate diagram of the tracking position of the four interactive targets in the method of the present invention. Detailed Implementation

[0050] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0051] It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0052] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0053] This application provides an embodiment of a multi-target interactive tracking method based on a gated neural network, such as... Figure 1 As shown, it includes the following steps:

[0054] Network design: includes two gating units containing neural network modules and a Kalman tracking framework;

[0055] The first-level gating unit, acting as a memory update module, is used to acquire the memory factor d, which represents the interaction information implicit in the historical states of multiple targets. k The second-level gating unit serves as an interactive input module, used for controlling the memory factor d. k The interactive information in the data is decoded and converted into a state space u. k Integrating the Kalman tracking framework;

[0056] The Kalman tracking framework includes a Kalman prediction model and a Kalman update model. The Kalman prediction model is used to determine the state space u. k The target state and covariance matrix are predicted; the Kalman update model updates the predicted target state and covariance matrix when it receives measurement data.

[0057] Network training: Using multi-objective historical state data, all learnable parameters in the network are trained by estimating the loss function between the current state and the true state;

[0058] Multi-target interactive tracking: Using a trained neural network to achieve multi-target interactive tracking.

[0059] This application's embodiments design a corresponding structure for each gating unit in a two-level gating framework, and utilize offline data to train the network, establishing corresponding mapping relationships to achieve stable and accurate target tracking. The designed network includes two gating units containing neural network modules, as well as a Kalman prediction and update part, such as... Figure 1As shown, this network is used to mine, memorize, and update interaction information in the historical states of multiple objectives, and decodes the memorized interaction information into acceleration control terms to participate in the prediction and update stages of the Kalman filter (KF). Based on KF, this network incorporates prior model information using the filtering formula and learns and trains the neural network using offline data. Each module has a clear objective and task, resulting in a simple structure.

[0060] Furthermore, the derivation process of the Kalman prediction model and the Kalman update model in this embodiment is as follows:

[0061] Step 1, Problem Modeling: Construct a GRNN based on the KF framework to effectively integrate offline data and prior model knowledge. This method comprehensively considers the evolution process of non-Markovian historical states and the interaction relationships of multiple objectives, aiming to achieve high-performance tracking in complex traffic scenarios.

[0062] When considering external input signals during multi-target motion, assuming there are N targets, the state prediction model of the i-th target can be described by the first-order Markov model shown in equation (1):

[0063]

[0064] in, f represents the target state at time k; k (·) is the state transition function; the control matrix G k Control input Transform into state space; noise distribution matrix Γ k Will have covariance The process noise is mapped onto the state space, representing its respective impact on the state components. This model exhibits the first-order Markov property, where each state is associated with its previous state, thus simplifying the state transition process and facilitating implementation and analysis. Furthermore, the recursive property allows for iterative updates to the state estimates.

[0065] Although the first-order Markov model is reasonable in many situations, in practical applications, the state of a target is not only affected by the state of the previous moment, but its state transition process is often more complex. Especially in complex traffic scenarios, it is also necessary to consider the interaction history information between multiple targets. Specifically, the relative speed, distance and motion intention between vehicle targets influence each other, forming complex trajectory correlations. At the same time, in specific scenarios, the driving interaction patterns of vehicles may exhibit fixed patterns. For example, at an intersection, the behavior of a vehicle is not only related to its own historical state, but also closely related to the behavior of other vehicles. This interaction information is implicit in the historical states of multiple targets, making it difficult to effectively model using the first-order Markov model. Therefore, the external input term u in equation (1) kIt can be viewed as an interactive variable influenced by the historical states of surrounding targets, and should be a non-Markov variable:

[0066]

[0067] Where, p u yes The output function extracts interaction information from the historical states of all interactive targets and transforms it into input for target state prediction.

[0068] Considering the variations in measurement data, the target measurement model can be described by equation (3):

[0069]

[0070] in, h is the measurement vector of the i-th target at time k; k (·) represents the measurement model; It has covariance Measurement noise; φ represents the empty set. It is worth noting that this invention assumes perfect data correlation to ensure that each detected measurement accurately corresponds to the target.

[0071] 1. Scenario (a): The system is able to acquire complete measurement data of the target;

[0072] 2. Scenario (b): Due to measurement fluctuations, measurement data for a specific target is continuously missing from a certain point in time.

[0073] Learning through GRNN This approach can incorporate complex interaction relationships into state estimation. These rich patterns are implicit in the offline data of multiple targets. The measurement data of multiple targets and their corresponding real states are assembled into usable offline data, as shown in (4):

[0074]

[0075] in, Represents the actual offline status value; k x,end Indicates the duration of state evolution (state-time sampling step); The measurement data of the target state in formula (3) in case (a) is used; m is the index of the multi-target trajectory dataset, with a total of M multi-target samples.

[0076] Modeling the state-space model in equations (1)-(3) while considering the dataset is challenging:

[0077] 1. Extracting hidden multi-objective interaction information from offline data is crucial to obtaining the known input term u in the state transition equation. k ;

[0078] 2. In the process of learning from offline data, it is crucial to properly integrate prior model knowledge;

[0079] To reduce computational complexity, an efficient non-Markov recursive estimation framework needs to be constructed.

[0080] 3. To reduce computational complexity, an efficient non-Markov recursive estimation framework needs to be constructed.

[0081] The ultimate goal is to build a KF-based GRNN that can effectively integrate offline data with prior model knowledge. This approach needs to consider the non-Markovian historical state evolution process and multi-objective interaction relationships to achieve high-performance tracking in complex traffic scenarios.

[0082] Step 2, KF-based tracking algorithm framework: Use GRNN to extract complex interaction relationships from the historical states of multi-target trajectories and transform them into multi-target interaction terms that are easy to integrate into the iterative update of KF. It includes the following main stages:

[0083] 1. Transform the non-Markov state-space model into an equivalent first-order Markov model with memory;

[0084] 2. Based on the KF framework, known input terms for learning by recurrent neural networks are introduced. Perform joint modeling;

[0085] Considering the non-Markovian variables in the equation, it is difficult to identify directly because it changes continuously with the accumulation of multiple target states, i.e., It is difficult to learn directly from multiple target historical states. Therefore, we attempt to break down the learning process into two steps:

[0086] 1. Introduce the memory factor d k It stores and continuously updates the interaction information implicit in the historical states of multiple targets.

[0087] 2. Regarding the memory factor d k Decode the interactive information in the middle and convert it into It is integrated into the Kalman tracking framework.

[0088] Therefore, it is described using a two-level nested function. Equation (5) can be obtained:

[0089]

[0090] in, It is d kThe output function encapsulates the implicit interaction information in the historical states of multiple objectives, representing the regular memory patterns in the interaction process; therefore, it is called interaction memory. Considering that the implicit interaction patterns at different times and in different objective states are similar, then d... k This can be approximated by nested functions as follows:

[0091]

[0092] At each moment, Interaction information extracted from the multi-objective historical state is selectively retained, while interaction information from the new moment is introduced. This nested form can be equivalently represented as:

[0093]

[0094] Non-Markov models can be equivalently represented by first-order Markov models by appropriately expanding the state dimension. That is, through equation (7), the non-Markov model described in (6) can be equivalently transformed into an iterable first-order Markov model with memory, as shown in the following equation:

[0095]

[0096] In this model, the recursion of the memory factor is influenced by the state, and the transition of the state is also influenced by the memory factor. If d k If we consider it as an extension of the state, then by x k and d k The u k It is still Markovian, and such a model preserves the interaction information of historical states while supporting recursive computation.

[0097] Based on the KF framework, interactive input terms learned by GRNN are introduced. If we consider a linear system, f in equations (8) and (3) k (·) and h k (·) can be derived from the state transition matrix F k and measurement matrix H k Alternative:

[0098]

[0099] Wherein, process noise w = [w x ,w y ] T ,satisfy Measurement noise v = [v px ,v py ,v vx ,v vy ] T ,satisfy and

[0100] Since the prediction and update models of KF used for each objective are exactly the same (the above derivation is only based on the i-th objective as an example), for the sake of simplicity, the prediction and update models of KF are omitted as follows:

[0101] Kalman prediction model:

[0102]

[0103] P k|k-1 =F k P k-1|k-1 (F k ) T +Q k (12)

[0104] Kalman update model:

[0105]

[0106] P k|k =(IK k H k )P k|k-1 (15)

[0107] in, Represents the state x at time k k The minimum mean square error posterior estimate; It can be considered as an acceleration input; P k|k-1 ,P k|k ,K k These represent the prediction covariance matrix, the state covariance matrix, and the Kalman gain, respectively.

[0108] When the measurement data is stable, filtering is performed using equations (11)-(15), and the result is denoted as FilteringOutput. When there is no measurement data input, filtering is performed using the Kalman prediction process of the prediction step in equations (11)-(12), and the result is denoted as Predicting Output. At this time, through network learning, stable state estimation can be maintained by utilizing the historical interaction information of multiple targets.

[0109] Furthermore, based on the Kalman framework described above, it is known that when using the Kalman prediction model, the state space u needs to be known. k Therefore, the above system uses a two-layer neural network with a hyperbolic tangent activation function to implement the state space u. k The prediction is as follows:

[0110] The memory update module and the interactive input module are represented as functions of equation (16):

[0111]

[0112] Among them, W d W u and b d ,b u d respectively k , The weights and biases, W d =(W d1 W d2 ), W u =(W u1 W u2 ), b d =(b d1 ,b d2 ), b u =(b u1 ,b u2 ), and W d W u ,b d ,b u ∈Ω, where Ω is the set of learnable parameters; σ and tanh are the sigmoid and hyperbolic tangent activation functions, respectively; i d i u These are the input vectors of the corresponding networks, and they satisfy:

[0113]

[0114] Where con{m,n} represents the concatenation function of vectors m and n; Ψ(·) is the maximum value normalization function; The filtered state of N interacting targets at time k-1; The difference vector of the filtered state of the target participating in the interaction at time k-1; Let N-1 be the filtered states of the N-1 targets participating in the interaction, excluding the i-th target, at time k-1, and satisfy the following:

[0115]

[0116] The splicing, normalization, and other operations performed are standard procedures in NN methods.

[0117] 1. For IMUG, its main function is to extract and store the interaction information contained in the historical states of multiple objectives in d. k In the process, the filtering iterative process is continuously updated. To reduce the learning difficulty of the network, in addition to providing the state information of multiple targets, the relative distance and relative velocity information between targets (i.e., differential state) are also provided. Input the NN module for learning and memorization.

[0118] 2. For IIG, its main function is to decode the interactive memory information and transform it into external input variables in the Kalman prediction process.

[0119] In equation (11), It can be viewed as an additional acceleration variable that corrects the target state, and is obtained by inputting the control term matrix Γ. k Attached to the target prediction state In the middle. The target predicted state is appended by the input control term matrix. Because d k It contains the state interaction information of the N targets under consideration, and the output is The term is information about the i-th objective. To reduce the learning burden on the network, the input to the network should be the state information of the other N-1 objectives besides the i-th objective, i.e., the input... For d k Decode to obtain

[0120] Compared to existing GRNNs (such as LSTM and GRU) combined with traditional filtering methods, the designed network is a low-level GRNN for state estimation based on the KF framework. Each gating unit has a specific function, and the design of each neural network module is targeted, reasonably interpretable, and reduces the need for learnable parameters and large amounts of training data to some extent.

[0121] Furthermore, the network designed in this embodiment is trained end-to-end under supervised learning by extracting latent interaction patterns from multi-objective historical state data. All learnable parameters in the network are trained by estimating the loss function between the current state and the ground truth state. This training method has been proven feasible.

[0122] Based on the minimum mean square error (MMSE) criterion commonly used in state estimation, the network is trained using a mean square error (MSE) loss function with L2 regularization. Considering the offline dataset D described in equation (4), the MSE loss function is defined as follows:

[0123]

[0124] Where, ω decay This is the scaling factor for adjusting the regularization term; i = {1, 2, ..., N} are the indices of the N interaction targets considered; m = {1, 2, ..., M} are the indices of the M training samples; k = {1, 2, ..., T} len / T} is a length of T len (s) is an index of time-sampling data.

[0125] Furthermore, mini-batch stochastic gradient descent is used to train Ω. Specifically, the entire dataset is divided into S (S < M) mini-batches, each with a batch size of b, and the index of each mini-batch is denoted as . At this point, the loss function for mini-batch is:

[0126]

[0127] Because the network has a recursive structure, the Backpropagation Time (BPTT) algorithm is used to train the network, which is a common method for training RNNs.

[0128] This application provides an embodiment of a multi-target interactive tracking device based on a gated neural network, comprising:

[0129] The memory update module is used to obtain the memory factor d, which can represent the interaction information implicit in the historical states of multiple objectives. k ;

[0130] An interactive input module is used for the memory factor d. k The interactive information in the data is decoded and converted into a state space u. k ;

[0131] Kalman prediction model, used to predict based on the state space u k Predict the target state and covariance matrix;

[0132] The Kalman update model updates the predicted target state and covariance matrix when measurement data is received.

[0133] Implementation Cases

[0134] The effects of this invention can be illustrated by the following experiments.

[0135] This invention evaluates the performance of the designed network using the real-world dataset NGSIM. The network was trained and tested using experimental data with varying numbers of interactive targets, and compared with model-based KF methods and advanced neural network-based MGBRNN methods. This comparison demonstrates the excellent state estimation performance of the designed network when measurement data fluctuates.

[0136] The experiment extracted trajectories with different numbers of interactive targets from the US-101 and I-80 highway datasets from NGSIM. Specifically, outliers were first removed, and then M groups (M=1000) of multi-target interactive trajectories were extracted for the numbers N of 3 and 4 interactive targets, with the difference between the horizontal and vertical distances within the range of [±3,±30] m. The frame period was T=0.1s, and the number of time samples was T. len / T = 100, state dimension is d x =4, and this {2,M,N,T len / T,d x The trajectory data of} is used as the training ground value. To meet experimental requirements, σ ​​was added to both the distance and velocity dimensions of each trajectory. p =0.5m and σ v Measurement noise of 0.2 m / s was used as the measurement data. And randomly select one target and delete its second half. The measurement data was used to simulate situations where the measurement data fluctuated. Of these data, 800 sets (80%) were randomly set up for testing, and 200 sets (20%) were used for verification and testing.

[0137] 1. Filtering parameters

[0138] The state vector is in and Let F represent the estimated values ​​of position and velocity in the x and y directions at time k, respectively. The measurements consist of noisy position and velocity data. State transition matrix F k It is a constant velocity model (CV model) with a sampling interval of T = 0.1s, and the measurement matrix H k For a linear model, F k and H k and the input control term matrix Γ k It can be written as:

[0139]

[0140] The process noise covariance matrix and the measurement noise covariance matrix are as follows:

[0141]

[0142] Where, q k =q a 2 I2,I d It is an identity matrix of dimension d. In the experiment, q a ,σ p ,σ v The values ​​were set to 2.5, 0.5 and 0.2 respectively.

[0143] 2. Training parameters

[0144] During training, the batch size was set to 50, the learning rate to 0.01, the interactive memory vector dimension to 64, and the training lasted for 50 epochs, using the Adam optimizer.

[0145] The experiment uses the relationship between the MSE (Mean Sequence Equation) of the network output filter state and the true trajectory value over time as the evaluation criterion. It is defined as follows:

[0146]

[0147] And their average value over the time dimension is defined as:

[0148]

[0149] The MSE represents the state and the position, respectively.

[0150] The experimental results are shown in Table 1. The tracking position coordinates for different target numbers are shown in the figure below. Figure 2 , 3 Figure 4 shows the performance of different methods when tracking sample trajectories with varying numbers of interacting targets. The MSE value indicates that trajectory stability decreases when measurement data fluctuates. Specifically, the KF method has the highest MSE for location information, exhibiting the lowest accuracy, especially in complex traffic scenarios where it fails to consider historical interactions between multiple targets, leading to a significant decrease in tracking accuracy when measurements fluctuate. In contrast, the EGBRNN method utilizes the time-related information of historical target trajectories, outperforming the KF method. However, this method relies excessively on SPG compensation for measurement fluctuations, which is still insufficient for high-precision requirements.

[0151] Table 1 Comparison of MSE (m) results from different methods

[0152]

[0153] The proposed method incorporates historical interaction data between multiple targets, demonstrating consistent superiority in tracking accuracy. This advantage is particularly evident when measurement data is missing, highlighting the robustness and reliability of the method in handling measurement fluctuations.

[0154] Furthermore, the designed GRNN-MTI is an efficient low-level model, and its concise gating structure reduces computational complexity. This design achieves excellent tracking performance while maintaining low complexity. Notably, even with a limited number of training samples, this GRNN with its small parameter size demonstrates significant advantages, making it highly suitable for practical traffic tracking tasks. Therefore, this algorithm has considerable engineering application value.

[0155] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-target interactive tracking method based on a gated neural network, characterized in that, Includes the following steps: Network design: includes two gating units containing neural network modules and a Kalman tracking framework; The first-level gating unit, acting as a memory update module, is used to acquire the memory factor d, which represents the interaction information implicit in the historical states of multiple targets. k The second-level gating unit serves as an interactive input module, used for controlling the memory factor d. k The interactive information in the data is decoded and converted into a state space u. k Integrating the Kalman tracking framework; The Kalman tracking framework includes a Kalman prediction model and a Kalman update model. The Kalman prediction model is used to determine the state space u. k Predict the target state and covariance matrix; The Kalman update model updates the predicted target state and covariance matrix when it receives measurement data. Network training: Using multi-objective historical state data, all learnable parameters in the network are trained by estimating the loss function between the current state and the true state; Multi-target interactive tracking: Using a trained neural network to obtain the target state, multi-target interactive tracking can be achieved.

2. The multi-target interactive tracking method based on a gated neural network according to claim 1, characterized in that, The first-level neural network serves as a memory update module: based on the interaction target at time k-1 in the filtered state. Difference vector and memory factor d k-1 Predict the memory factor d at time k k .

3. The multi-target interactive tracking method based on a gated neural network according to claim 2, characterized in that, The memory factor d k for: d k =W d2 (σ(W d1 i d +b d1 ))+b d2 Among them, W d and b d d respectively k The weights and biases, W d =(W d1 W d2 ) and b d =(b d1 ,b d2 ) represents the set of learnable parameters, σ() is the hyperbolic tangent activation function, con{} represents the concatenation function, and Ψ(·) is the maximum normalization function.

4. The multi-target interactive tracking method based on a gated neural network according to claim 2, characterized in that, The second-level neural network serves as the interactive input module: based on the filtered states of the N-1 interactive targets (excluding the i-th target) at time k-1. and memory factor d k Predict the state space u at time k k .

5. The multi-target interactive tracking method based on a gated neural network according to claim 4, characterized in that, The state space u k Among them, W u and b u for The weights and biases, W u =(W u1 W u2 ) and b u =(b u1 ,b u2 ) represents the set of learnable parameters, tanh() is the hyperbolic tangent activation function, con{} represents the concatenation function, and Ψ(·) is the maximum normalization function.

6. The multi-target interactive tracking method based on a gated neural network according to claim 1, characterized in that, The Kalman prediction model is as follows: P k|k-1 =F k P k-1|k-1 (F k ) T +Q k in, This represents the predicted value of the target state at time k. F represents the predicted value of the target state at time k-1. k Let G represent the state transition matrix. k P represents the control matrix. k|k-1 Let P represent the predicted covariance matrix at time k. k-1|k-1 Let Q represent the predicted covariance matrix at time k-1. k The covariance matrix represents the process noise.

7. The multi-target interactive tracking method based on a gated neural network according to claim 6, characterized in that, The Kalman update model is as follows: P k|k =(I-K k H k )P k|k-1 Among them, H k Measurement matrix, R k Indicates measurement noise, z k K represents the measurement data. k This represents the Kalman gain.

8. The multi-target interactive tracking method based on a gated neural network according to claim 4, characterized in that, The minimum mean square error (MSE) criterion based on state estimation is used to train the network by constructing a loss function with L2 regularization of the MSE. The MSE loss function is defined as follows: Where, ω decay This is the scaling factor for adjusting the regularization term, where i = {1, 2, ..., N} are the indices of the N interaction targets, m = {1, 2, ..., M} are the indices of the M training samples, and k = {1, 2, ..., T} are the values ​​of the regularization term. len / T} is a length of T len (s) is the index of the time-sampling data, where Ω is the set of learnable parameters.

9. The multi-target interactive tracking method based on a gated neural network according to claim 8, characterized in that, Mini-batch stochastic gradient descent is used to train and update Ω.

10. A multi-target interactive tracking device based on a gated neural network, characterized in that, include: The memory update module is used to obtain the memory factor d, which can represent the interaction information implicit in the historical states of multiple objectives. k ; An interactive input module is used for the memory factor d. k The interactive information in the data is decoded and converted into a state space u. k ; Kalman prediction model, used to predict based on the state space u k Predict the target state and covariance matrix; The Kalman update model updates the predicted target state and covariance matrix when measurement data is received.

Images