A multi-target tracking filtering method and device

Abstract

The present invention discloses a multi-target tracking filtering method and device, comprising the following steps: obtaining measurement data of a target, inputting the measurement data into a trained multi-target tracking model, extracting features from the measurement data based on an encoder, and generating an improved representation of the measurement data; cascading tracking queries and detection queries into a decoder, processing the improved representation of the measurement data, and obtaining a query vector; performing soft data association on the query vector based on the decoder to obtain a decoding result; inputting the decoding result into a predictor and a classifier for processing, and generating a state prediction result for each time window; processing the state prediction results of all time windows under a cross-frame target alignment mechanism based on a QTM module, obtaining multi-target tracking query results for all time windows, and generating target motion trajectories. The multi-target tracking filtering method of the present invention improves the long-term tracking accuracy of multiple targets and can generate predicted trajectories.

Description

Technical Field

[0001] The invention belongs to the technical field of radar target tracking, and in particular relates to a multi-target tracking filtering method and device. Background Art

[0002] Multi-Target Tracking (MTT) is a key task of modern radar systems. It aims to estimate the motion states of multiple targets, such as position and velocity, in real time based on radar observation information, and to track their motion trajectories continuously and accurately. As one of the core issues in the field of radar signal processing, MTT has extensive and important applications in many fields such as military, autonomous driving, and modern urban monitoring. However, due to the complexity and variability of radar detection environments, such as low signal-to-noise ratio, high clutter density, and frequent target appearance and disappearance, MTT has always faced severe technical challenges.

[0003] Traditional MTT methods are mainly based on the Random Finite Set (RFS) statistical framework, which models multi-target states and observations as random sets and uses the Bayesian filtering mechanism to achieve recursive state estimation. Under the guidance of RFS theory, researchers have developed a series of classic MTT filters, including probability hypothesis density (PHD) filters, generalized labelled multi-Bernoulli (GLMB) filters, and so on. Although these methods have achieved gratifying results in a variety of scenarios, they generally rely on prior models of target motion and observation noise, such as Gaussian assumptions and Poisson distributions, and are difficult to adapt to model mismatch problems in complex dynamic environments. In addition, these filters also have certain limitations in terms of real-time performance and computational efficiency.

[0004] In recent years, deep learning, with its powerful feature extraction and function fitting capabilities, has brought new opportunities to solve the MTT problem. Through deep neural networks, the intrinsic patterns of target motion and observation are mined from data, reducing the reliance on prior assumptions and improving the generalization performance of the algorithm. Existing work can be divided into two categories: one representative work combines deep learning with the classic MTT filter, and uses neural networks to improve the key modules of the filter under the original filtering framework, such as target detection, clutter suppression, state estimation, etc.; the other attempt explores the MTT paradigm of deep network models and realizes the simultaneous estimation of multi-target states and trajectories through radar observation information. Although these methods have achieved better performance than traditional methods in specific scenarios, it is still difficult to achieve accurate characterization and long-term tracking of target motion when faced with challenges such as the inherent sparsity of radar data and the complexity of data association.

[0005] Recently, Transformer, with its flexible attention mechanism and powerful time series modeling capabilities, has provided a new idea for further breaking through the bottleneck of MTT. Pinto et al. proposed a model framework Multi-Target Tracking Transformer (MT3) for multi-target tracking based on the Transformer network. With the unique self-attention and cross-attention mechanisms of Transformer, MT3 can learn the target motion pattern from the measurement information and establish the long-range dependency of the target sequence, implicitly perform data association, and realize end-to-end multi-target tracking. The performance on a variety of complex tasks exceeds that of the traditional model-based MTT method. However, the design of the MT3 model is to use all radar measurement data in a time period to predict the state and existence probability of all targets in the last time step. It is only a single time step prediction algorithm and cannot characterize the long-term tracking process of multiple targets and generate tracking trajectories.

[0006] Therefore, it is urgent to propose a multi-target tracking filtering method and device to solve the problems of data association and long-term accurate target tracking and generate target motion trajectories. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a multi-target tracking filtering method and device to solve the problems existing in the above prior art.

[0008] To achieve the above object, the present invention provides a multi-target tracking filtering method, comprising the following steps:

[0009] Obtaining measurement data of the target;

[0010] Inputting the measurement data into a trained multi-target tracking model, performing feature extraction on the measurement data based on an encoder, and generating an improved representation of the measurement data;

[0011] The tracking query and the detection query are cascaded and input into a decoder, and the improved representation of the measurement data is processed to obtain a query vector;

[0012] Performing data association on the query vector based on the decoder to obtain a decoding result;

[0013] Inputting the decoding results into a predictor and a classifier for processing to generate a state prediction result for each time window;

[0014] Based on the QTM module, according to the state prediction results of each time window, the multi-target tracking query results of all time windows are obtained, and the target motion trajectory is generated.

[0015] Optionally, the process of obtaining measurement data of the target includes: obtaining motion state information of the target based on the radar, obtaining measurement information of the target based on the motion state information, and combining the measurement information of the target with all clutter measurements at a corresponding time step to obtain final measurement data.

[0016] Optionally, the parameters of the encoder are set to: input sequence dimension 256; position encoding dimension 256; output sequence dimension 256; number of multi-head attention channels 8; feedforward neural network hidden layer dimension 2048; neuron drop probability 0.1; activation function ReLU; number of encoder layers 6.

[0017] Optionally, the tracking query and the detection query are cascaded and input into the decoder, and the process of processing the improved representation of the measurement data includes: constructing a tracking query based on the number of tracking queries and the embedding dimension of the query, and constructing a detection query based on the number of detection queries and the embedding dimension of the query; predicting the status of an existing target based on the interaction between the tracking query and the improved representation of the measurement data; detecting a newly appeared target based on the interaction between the detection query and the improved representation of the measurement data; and generating a query vector based on the query results of the tracking query and the detection query.

[0018] Optionally, the process of performing soft data association on the query vector based on the decoder includes: the decoder implicitly performs soft data association on the query vector through self-attention and cross-attention mechanisms.

[0019] Optionally, the parameters of the decoder are set as follows: the decoder parameters are set as follows: input sequence dimension 256; position encoding dimension 256; output sequence dimension 256; number of multi-head attention channels 8; hidden layer dimension of the feedforward neural network 2048; neuron drop probability 0.1; activation function ReLU; number of decoder layers 6.

[0020] Optionally, the parameters of the predictor are set as follows: input dimension 256; hidden layer dimension 128; output layer dimension 2; multi-layer perceptron layer number 3;

[0021] The parameters of the classifier are set as: input dimension 256; output dimension 1.

[0022] Optionally, the process of processing the state prediction results of all time windows under the cross-frame target alignment mechanism based on the QTM module to obtain the multi-target tracking query results of all time windows includes: mapping the state prediction results of the current frame based on a multi-layer perceptron to obtain a hidden state, and dividing the hidden state into a detection query hidden state and a trajectory query hidden state; obtaining a screened detection query hidden state and a trajectory query hidden state after confidence screening of the detection query hidden state and the trajectory query hidden state; adding the tracking query result of the previous frame and the screened trajectory query hidden state of the current frame as the query and key value of the multi-head autonomous intention, and splicing the addition result after encoding processing and the screened detection query hidden state of the current frame to obtain the tracking query result of the next frame.

[0023] The present invention also provides a multi-target tracking filtering device for implementing a multi-target tracking filtering method, comprising: a data acquisition module, an encoder, a data query module, a decoder, a predictor, a classifier and a QTM module;

[0024] The data acquisition module is used to acquire measurement data of the target;

[0025] The encoder is used to extract features from the measurement data input into the trained multi-target tracking model to generate an improved representation of the measurement data;

[0026] The data query module is used to cascade input tracking query and detection query in the decoder, process the improved representation of the measurement data generated by the encoder, and obtain a query vector;

[0027] The decoder is used to perform soft data association on the query vector output by the data query module to obtain a decoding result;

[0028] The predictor and the classifier are used to process the decoding results and generate a state prediction result for each time window;

[0029] The QTM module is used to process the state prediction results of all time windows generated by the data prediction module, obtain the multi-target tracking query results of all time windows, and generate the target motion trajectory.

[0030] Optionally, the data query module includes a tracking query unit and a detection query unit;

[0031] The tracking query unit is used to construct a tracking query based on the number of tracking queries and the embedding dimension of the query, and to interact with the improved representation of the measurement data based on the tracking query to predict the state of the existing target;

[0032] The detection query unit is used to construct a detection query based on the number of detection queries and the embedding dimension of the query, and to detect a newly appeared target based on the interaction between the detection query and the improved representation of the measurement data.

[0033] Compared with the prior art, the present invention has the following advantages and technical effects:

[0034] The present invention cascades tracking queries and detection queries into a decoder, which are used to update the state of known targets and detect new targets, respectively, to obtain query vectors; then, based on the decoder, soft data association is performed on the query vectors to obtain decoding results, which further enhances the tracking query's ability to describe target motion, adaptively aggregates target motion information at different time scales, and greatly improves the robustness of the tracker;

[0035] The present invention inputs the decoding results into the predictor and classifier for processing, and generates the state prediction results of each time window; based on the QTM module, the state prediction results of all time windows are processed under the cross-frame target alignment mechanism, and the multi-target tracking query results of all time windows are obtained to generate the target motion trajectory. Among them, the QTM module realizes the long-term association between the tracking query and the real trajectory by transmitting matching information between adjacent frames, greatly enhancing the temporal continuity and consistency of the tracking process, improving the multi-target long-term tracking accuracy, and facilitating the generation of predicted trajectories. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 is a flow chart of a multi-target tracking method according to an embodiment of the present invention;

[0038] Figure 2 A network model structure diagram of a multi-target tracking method according to an embodiment of the present invention;

[0039] Figure 3 It is a QTM module structure diagram of the multi-target tracking method according to an embodiment of the present invention;

[0040] Figure 4 Target trajectory estimation diagrams of a multi-target tracking method according to an embodiment of the present invention, wherein (a) is a target trajectory estimation diagram of a multi-target tracking method according to scenario one, (b) is a target trajectory estimation diagram of a multi-target tracking method according to scenario two, and (c) is a target trajectory estimation diagram of a multi-target tracking method according to scenario three;

[0041] Figure 5 A comparison chart of the GOSPA evaluation indexes of the multi-target tracking method according to an embodiment of the present invention and other benchmark methods;

[0042] Figure 6 A comparison chart of the Pro-GOSPA evaluation indicators of the multi-target tracking method according to an embodiment of the present invention and other benchmark methods. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] like Figure 1 As shown, this embodiment provides a multi-target tracking filtering method, including the following steps:

[0046] Acquire measurement data of the target, input the measurement data into a trained multi-target tracking model (Track-MT3 model), perform feature extraction on the measurement data based on the encoder, and generate an improved representation of the measurement data;

[0047] The tracking query and the detection query are cascaded and input into a decoder, and the improved representation of the measurement data is processed to obtain a query vector;

[0048] Performing soft data association on the query vector based on a decoder to obtain a decoding result;

[0049] Inputting the decoding results into a predictor and a classifier for processing to generate a state prediction result for each time window;

[0050] Based on the QTM module, the state prediction results of all time windows are processed under the cross-frame target alignment mechanism, the multi-target tracking query results of all time windows are obtained, and the target motion trajectory is generated.

[0051] As a specific embodiment, the following steps are included:

[0052] Step 1: Before inputting the measurement data into the trained Track-MT3 model, the trained Track-MT3 model weights need to be loaded;

[0053] The multi-target tracking method provided by the present invention does not rely on a priori models of target motion and measurement noise, but builds a deep model to learn the target's motion pattern and the inherent characteristics of measurement information from massive training data end-to-end, thereby realizing long-term multi-target tracking. Before using the model, it is necessary to train the model according to a certain strategy, load the trained model weights when using it, and then infer the tracking results based on the measurement information.

[0054] Step 2: The process of obtaining the target's measurement data includes: obtaining the target's motion state information based on the radar, obtaining the target's measurement information based on the motion state information, combining the target's measurement information with all clutter measurements at the corresponding time step, and obtaining the final measurement data.

[0055] As a specific embodiment, the working principle of radar is to measure the distance and direction of the target by reflecting electromagnetic waves. The radar sends out a specific signal, and the signal is reflected back after reaching the target. The reflected signal is processed by the signal processing device and then transmitted to the filtering device. In the filtering device, information such as the speed, heading and shape of the target is estimated. The measurement information obtained by the radar generally contains noise pollution. First, the target's motion state information and measurement information are modeled as follows:

[0056]

[0057] Where F is the state transfer matrix, is the process noise, which has a mean of zero and a covariance matrix of Q multivariate Gaussian distribution. H is the measurement matrix, is the measurement noise, which follows a Gaussian distribution with zero mean and covariance matrix R.

[0058] In addition, due to the performance and characteristics of the sensor itself, the measurement information obtained by the filter contains not only the real measurement from the target, but also the interference information of the clutter. Assume that at time step k, the number of clutter measurements n c (k) is subject to the parameter λ c Poisson distribution n c (k)~Poission(λ c ),in The average value of the clutter measurement. The location of the clutter (x c ,y c ) in the sensor’s field of view (FOV x ,FOV y )∈R 2 The interior follows a uniform distribution:

[0059] x c ~U(x lb ,x ub )

[0060] yc ~U(y lb ,y ub )

[0061] where x lb 、x ub and lb ,y ub are the lower and upper bounds of the x-axis and y-axis of the field of view respectively. Finally, all the measured values ​​Z at time step k are k The collection is:

[0062]

[0063] Among them C k are all clutter measurements at this time step.

[0064] Step 3: extracting features from the measurement data based on the encoder to generate an improved representation of the measurement data;

[0065] As a specific example, assume that the measurement data obtained within a time period τ is represented by z 1:n =[Z T-τ ,…,Z T ], every 20 sampling moments is a sliding window. Figure 2 As shown, first, the Transformer encoder processes the measurement sequence z 1:n , generating an improved representation e for each time window measurement i , which can encode useful attributes. The encoder consists of N layers with the same function f i enc (·), i = 1,…, N are stacked, and each layer contains two sub-layers: a multi-head self-attention layer MultiHead(·) and a feed-forward layer FFN(·), as well as a residual connection and layer normalization LayerNorm(·) between the two sub-layers.

[0066] The specific parameter settings of the encoder are as follows: input sequence dimension: 256; position encoding dimension: 256; output sequence dimension: 256; number of multi-head attention channels: 8; feedforward neural network hidden layer dimension: 2048; neuron drop probability: 0.1; activation function: ReLU; number of encoder layers: 6.

[0067] Assume the input of the l-th layer encoder is x (l-1) , the output is x (l) , then the calculation process of the encoder can be expressed as:

[0068] a (l) =MultiHead(x (l-1) )

[0069]

[0070] Among them, x (0) =Ez+p, is the measurement embedding matrix, is the position code, the encoder output e 1:n =x (N) .

[0071] Step 4: Concatenate the tracking query and the detection query into a decoder, process the improved representation of the measurement data, and obtain a query vector;

[0072] As a specific embodiment, in order to adapt to the change of the number of targets in the MTT task, such as Figure 2 As shown, Track-MT3 introduces two types of queries: detection queries and tracking query They are responsible for detecting new targets and predicting the status of existing targets, respectively. d ,N tr are the number of detection queries and tracking queries, respectively, and d q is the embedding dimension of the query. The tracking query and the detection query are cascaded into the decoder and the encoder output e i to interact.

[0073] Unlike detection queries, which have a fixed set size, tracking queries are dynamically updated sets with variable sizes, used to track already-appeared targets. The detection query size is fixed at 16, and the number of tracking queries is initially 0, which is then adaptively adjusted based on the detection results.

[0074] Step 5: Performing soft data association on the query vector based on the decoder to obtain a decoding result;

[0075] As a specific embodiment, in step 3, an improved representation of the measurement data is obtained. i , and the concatenated query vector [q d ;q tr ], based on the encoding representation e 1:n , the decoder implicitly performs soft data association through self-attention and cross-attention mechanisms to calculate the decoding result The decoder stacks M layers with the same function f i dec (·), i = 1, ..., M. In addition to the two sub-layers in the encoder, each layer also introduces a third sub-layer CrossMultiHead (·) for interacting with the encoder output.

[0076] The specific parameter settings are as follows: input sequence dimension: 256; position encoding dimension: 256; output sequence dimension: 256; number of multi-head attention channels: 8; hidden layer dimension of feedforward neural network: 2048; neuron drop probability: 0.1; activation function: ReLU; number of decoder layers: 6.

[0077] Assume the object query matrix is The input of the l-th layer decoder is y (l-1) , the calculation process of the decoder is:

[0078] q (l) =MultiHead(y (l-1) )

[0079]

[0080] Step 6: Input the decoding result into the predictor and classifier for processing to generate a state prediction result for each time window;

[0081] As a specific embodiment, the decoder output in step 5 Input predictor and classifier, predictor is a multi-layer perceptron (MLP), classifier is composed of a linear layer (FC). After processing by predictor and classifier, output state prediction sequence in Represents the target state of the i-th prediction and the existence probability p i , n is the maximum number of potential targets.

[0082] The specific parameters of the predictor and classifier are set as follows:

[0083] Predictor: Input dimension: 256; Hidden layer dimension: 128; Output layer dimension: 2; Multilayer perceptron layer number: 3;

[0084] Classifier: Input dimension: 256; Output dimension: 1;

[0085] The predictor and classifier calculation process can be expressed as:

[0086] Step 7: Based on the QTM module, the state prediction results of all time windows are processed under the cross-frame target alignment mechanism to obtain the multi-target tracking query results of all time windows and generate the target motion trajectory.

[0087] As a specific example, the effectiveness of tracking query depends on its ability to encode the target state. To further enhance the modeling of target motion by tracking query, Track-MT3 has designed a QTM module, such as Figure 3After the calculations of steps 3, 4, 5, and 6 for each time window, the QTM module is used in this step to predict the target state of the previous moment calculated in step 6. As input, output the tracking query of the i-th frame The calculation process of the QTM module is as follows:

[0088]

[0089] Among them, MLP is a multi-layer perceptron. To query the prediction result after the decoder and feature interaction in the i-1th frame, the hidden state H is obtained after zero-bias multi-layer perceptron mapping i-1 , the Split operation divides the hidden state into the hidden states of the detection query and the trajectory query Then, after confidence screening, the hidden state of the new target is obtained, and the hidden state of the extinct target is eliminated, and we get and The TFN module is an improved Transformer encoder that converts the tracking query of the previous frame As additional information and the current frame's filtered trajectory query hidden state Add as the query and key value of multi-head autonomous intention, after being processed by this encoder and the current frame filtered detection query hidden state Splice to get the tracking query of the i-th frame

[0090] The embodiment of the present invention also provides a cross-frame target alignment mechanism for a multi-target tracking algorithm. In the QTM module of step 7, in order to ensure the consistent correspondence between the tracking query and the target, that is, each tracking query focuses on tracking a specific target, Track-MT3 introduces a CTA mechanism: "target consistency matching" for tracking queries, and "new target matching" for detection queries.

[0091] For tracking queries:

[0092] make Indicates J in the i-th frame i The predicted results of the tracking query, is the binary matching matrix between tracking query and truth value, where M i is the number of true value targets. When i>1, the label assignment of tracking query follows the following principle:

[0093]

[0094] That is, the supervision information of the tracking query comes from the matching result of the previous frame (i-1) and newly detected target matches This strategy implicitly adjusts the number of tracking queries to adapt to the changes in the number of targets, i.e. For the first frame (i=1), due to the lack of prior tracking information, Is an empty set.

[0095] For detection queries:

[0096] set up is the prediction of K detection queries in the i-th frame, is the set of newly appeared true value targets, then the optimal detection match Should meet:

[0097]

[0098] Where C(·) is the matching cost function, and Ω is the set of all feasible bipartite matching matrices. In Track-MT3, a cost function that comprehensively considers state consistency and category consistency is designed:

[0099]

[0100] in is the cost between the kth detection result and the mth new target, defined as:

[0101]

[0102] in For detection status and truth state The absolute distance between For detection category and truth value categories The two terms measure the difference between the detection result and the true value target in terms of state and category:

[0103]

[0104] Where N c =2 is the number of categories, 1 is the indicator function, is the predicted probability that the kth detection result belongs to the jth class. The bipartite matching problem based on this cost function can be solved by the Hungarian algorithm.

[0105] On the other hand, based on the same inventive concept as the above embodiment, this embodiment further provides a multi-target tracking filtering device, and the use method of the multi-target tracking filtering device and the multi-target tracking filtering method provided by the above embodiment can be mutually referenced in effect. The multi-target tracking filtering device includes: a data acquisition module, a feature extraction module, a data query module, a data association module, a data prediction module and a result generation module;

[0106] The data acquisition module is used to acquire measurement data of the target;

[0107] The feature extraction module is used to input the measurement data acquired by the data acquisition module into the trained Track-MT3 model, perform feature extraction on the measurement data based on the encoder, and generate an improved representation of the measurement data;

[0108] The data query module is used to cascade input tracking query and detection query in the decoder, process the improved representation of measurement data generated by the feature extraction module, and obtain a query vector;

[0109] The data association module is used to perform soft data association on the query vector output by the data query module based on the decoder to obtain a decoding result;

[0110] The data prediction module is used to input the decoding results generated by the data association module into the predictor and classifier for processing, and generate the state prediction results of each time window;

[0111] The result generation module is used to process the state prediction results of all time windows generated by the data prediction module under the cross-frame target alignment mechanism based on the QTM module, obtain the multi-target tracking query results of all time windows, and generate the target motion trajectory.

[0112] As a specific embodiment, the data query module includes a tracking query unit and a detection query unit; the tracking query unit is used to construct a tracking query based on the number of tracking queries and the embedding dimension of the query, and interact with the improved representation of the measurement data based on the tracking query to predict the status of an existing target; the detection query unit is used to construct a detection query based on the number of detection queries and the embedding dimension of the query, and interact with the improved representation of the measurement data based on the detection query to detect a newly appeared target.

[0113] The effects of the present invention are further described below in conjunction with simulation experiments and accompanying drawings.

[0114] Simulation conditions of the present invention:

[0115] System: 64-bit Windows 11 operating system;

[0116] CPU: 12th Gen Intel(R)Core(TM)i5-124002.50GHz;

[0117] GPU: NVIDIA GeForce RTX 3090Ti;

[0118] Python: 3.7.4;

[0119] Pytorch: 1.6.0;

[0120] Torchvision:0.7.0.

[0121] A multi-target motion scene is constructed, GOSPA and Pro-GOSAP are used as performance evaluation indicators, and the experimental results are analyzed through simulation experiments.

[0122] In the Cartesian coordinate system, the radar measurement field of view is set to 20×20 (m 2 ), the target moves at a uniform speed in the field of view, and the state transfer matrix is:

[0123]

[0124] The process noise covariance matrix Q is:

[0125]

[0126] Where Δt = 0.1 (s) represents the time interval between two consecutive observations, σ q is the process noise intensity. The mean μ of the target initial position (x, y) (x,y) =

[00] T and variance Target initial speed (v x ,v y ) and variance

[0127] The sensor measurement equation is:

[0128]

[0129] The measurement noise covariance is:

[0130]

[0131] where σ r To measure the noise intensity.

[0132] Three different scenarios were designed to simulate the target movement under 80 time steps: Scenario 1 is a simple tracking scenario, with a small number of targets (6) and small dynamic changes of targets (birth rate 0.04, death rate 0.01). Under this ideal condition, the number of targets is small and the duration of the motion trajectory is long, which examines the basic tracking ability of Track-MT3; Scenario 2 increases the dynamic changes of the targets on the basis of Scenario 1 (birth rate 0.08, death rate 0.02). Under this condition, the number of targets remains unchanged, but the duration of the target trajectory is short, which examines the algorithm's ability to detect targets; Scenario 3 further increases the number of targets (10), and the birth rate (0.12) and death rate (0.03) of the targets are also higher than the previous two scenarios. In general, Scenario 3 has the greatest tracking difficulty, and requires higher target detection and long-term tracking capabilities of the multi-target tracking algorithm.

[0133] Figure 4 (a)-(c) show the tracking results of Track-MT3 algorithm in three scenarios. The solid lines connecting the hollow circles in the figure represent the real trajectory of the target at each moment, the solid lines connecting the cross symbols represent the estimated trajectory of the target, the triangle represents the starting point of the target, the quadrilateral represents the end point of the target, and the dot symbol represents clutter and false measurement. In scenario 1, P is set D =0.95,σ q =0,σ r =0,λ c =5, it can be seen that Track-MT3 can track all targets accurately and stably; in scenario 2, P is set D =0.9,σ q =0.04,σ r =0.02, λ c = 10, although a small number of false alarm tracks appear, the overall tracking effect is still good; while in scenario 3, P is set D =0.85,σ q =0.08,σ r =0.03,λ c =15. In this complex environment with high noise, high false alarm and high target density, Track-MT3 can still complete target detection and trajectory generation well, which reflects its long-term tracking capability in complex dynamic environment.

[0134] In addition, different detection probabilities P are set in scenario 3. D , process noise intensity σ q , measurement noise intensity σ r and the number of clutter λ c To change the data association difficulty and tracking complexity, verify the robustness of the algorithm: Experiment 1 (P D =0.95,σ q =0,σr =0,λ c =5); Experiment 2 (P D =0.9,σ q =0.04,σ r =0.02, λ c =10); Experiment 3 (P D =0.95,σ q =0,σ r =0,λ c =5). We selected three benchmark methods, Method 1 (Random Prediction), Method 2 (True Measurement at the current moment), and Method 3 (MT3 model) as references, and compared the performance evaluation indicators of GOSPA and Pro-GOSPA through 50 Monte Carlo simulation calculations.

[0135] Depend on Figure 5 From the GOSPA indicator line chart, the MT3 algorithm has the best performance at the initial moment. The evaluation index is higher than the random prediction between the 10th and 20th time steps, and then gradually stabilizes and the trend of convergence over time is not obvious. The Track-MT3 algorithm is slightly higher than MT3 at the initial time, but can converge quickly over time. It is better than MT3 in the 5th to 10th time steps, and the convergence trend is obvious. As the tracking complexity increases, Track-MT3 is better than the MT3 algorithm, indicating that it can better adapt to complex tracking environments.

[0136] like Figure 6 As shown in the figure, Pro-GOSPA is slightly lower than GOSPA, and the change trend of Pro-GOSPA indicators is different from that of GOSPA. MT3 is better than Track-MT3 in the initial time, but tends to diverge over time. This is because MT3 extracts features from the target's historical state information and cannot characterize the target's motion for a long time, resulting in a decrease in the confidence of the target's existence probability prediction. The Pro-GOSPA indicators of Track-MT3 in the three experiments are significantly better than MT3, and the convergence trend is obvious, especially in Experiment 1. This shows that Track-MT3 can more accurately judge the authenticity of the target, reduce the impact of false alarms, and track the target for a long time.

[0137] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A multi-target tracking filtering method, characterized in that: The following steps are involved: Obtaining measurement data of the target; Inputting the measurement data into a trained multi-target tracking model, performing feature extraction on the measurement data based on an encoder, and generating an improved representation of the measurement data; The tracking query and the detection query are cascaded and input into a decoder, and the improved representation of the measurement data is processed to obtain a query vector; Performing soft data association on the query vector based on a decoder to obtain a decoding result; Inputting the decoding results into a predictor and a classifier for processing to generate a state prediction result for each time window; Based on the QTM module, according to the state prediction results of each time window, the multi-target tracking query results of all time windows are obtained, and the target motion trajectory is generated; The process of processing the state prediction results of all time windows under the cross-frame target alignment mechanism based on the QTM module to obtain the multi-target tracking query results of all time windows includes: obtaining the hidden state after mapping the state prediction results of the current frame based on the multi-layer perceptron, and dividing the hidden state into a detection query hidden state and a trajectory query hidden state; obtaining the filtered detection query hidden state and trajectory query hidden state after confidence screening of the detection query hidden state and the trajectory query hidden state; adding the tracking query result of the previous frame and the filtered trajectory query hidden state of the current frame as the query and key value of the multi-head self-attention, and splicing the addition result after encoding processing and the filtered detection query hidden state of the current frame to obtain the tracking query result of the next frame.

2. The multi-target tracking filtering method according to claim 1, characterized in that: The process of obtaining the measurement data of the target includes: obtaining the motion state information of the target based on the radar, obtaining the measurement information of the target based on the motion state information, combining the measurement information of the target with all the clutter measurements at the corresponding time step, and obtaining the final measurement data.

3. The multi-target tracking filtering method according to claim 1, characterized in that: The tracking query and the detection query are cascaded and input into the decoder, and the process of processing the improved representation of the measurement data includes: constructing a tracking query based on the number of tracking queries and the embedding dimension of the query, and constructing a detection query based on the number of detection queries and the embedding dimension of the query; predicting the state of an existing target based on the interaction between the tracking query and the improved representation of the measurement data; detecting a newly appeared target based on the interaction between the detection query and the improved representation of the measurement data; and generating a query vector based on the query results of the tracking query and the detection query.

4. The multi-target tracking filtering method according to claim 1, characterized in that: The process of performing soft data association on the query vector based on the decoder includes: the decoder implicitly performs soft data association on the query vector through self-attention and cross-attention mechanisms.

5. The multi-target tracking filtering method according to claim 1, characterized in that: The parameter settings of the encoder and decoder are: The decoder parameters are set as follows: input sequence dimension 256; position encoding dimension 256; output sequence dimension 256; multi-head attention channel number 8; hidden layer dimension of feedforward neural network 2048; neuron drop probability 0.1; activation function ; The number of decoder layers is 6.

6. The multi-target tracking filtering method according to claim 1, characterized in that: The parameters of the predictor are set as follows: input dimension 256; hidden layer dimension 128; output layer dimension 2; and multi-layer perceptron layer number 3.

7. The multi-target tracking filtering method according to claim 1, characterized in that: The parameters of the classifier are set as: input dimension 256; output dimension 1.

8. A multi-target tracking filtering device, used to implement the multi-target tracking filtering method according to any one of claims 1 to 7, characterized in that: include: Data acquisition module, encoder, data query module, decoder, predictor, classifier and QTM module; The data acquisition module is used to acquire measurement data of the target; The encoder is used to extract features from the measurement data input into the trained multi-target tracking model to generate an improved representation of the measurement data; The data query module is used to cascade input tracking query and detection query in the decoder, process the improved representation of the measurement data generated by the encoder, and obtain a query vector; The decoder is used to perform soft data association on the query vector output by the data query module to obtain a decoding result; The predictor and the classifier are used to process the decoding results and generate a state prediction result for each time window; The QTM module is used to process the state prediction results of all time windows generated by the data prediction module, obtain the multi-target tracking query results of all time windows, and generate the target motion trajectory.

9. The multi-target tracking filtering device according to claim 8, characterized in that: The data query module includes a tracking query unit and a detection query unit; The tracking query unit is used to construct a tracking query based on the number of tracking queries and the embedding dimension of the query, and to interact with the improved representation of the measurement data based on the tracking query to predict the state of the existing target; The detection query unit is used to construct a detection query based on the number of detection queries and the embedding dimension of the query, and to detect a newly appeared target based on the interaction between the detection query and the improved representation of the measurement data.

Images