Methods for generating multi-target tracking models and multi-target tracking methods in cluttered environments

Abstract

This invention discloses a method for generating a multi-target tracking model and a multi-target tracking method in cluttered environments, comprising: acquiring training data of a multi-target scene; constructing a radar multi-target tracking network based on a cross-attention mechanism; and training the radar multi-target tracking network using the training data. The multi-target tracking model proposed in this invention can learn the data association probability between the estimated state of the target at the previous moment and the measurement value at the current moment through a cross-attention mechanism layer without any prior information, thereby obtaining the target measurement information at the current moment after clutter removal; furthermore, it can use a GRU network to filter the sum of target measurement-related information obtained through residual connections to obtain the estimated state of the target at the current moment, achieving multi-target tracking; compared with existing technologies, using the trained multi-target tracking model for multi-target tracking tasks does not require adjusting any parameter values, and the tracking effect is better.

Description

Technical Field

[0001] This invention belongs to the field of radar target tracking technology, specifically relating to a method for generating a multi-target tracking model and a multi-target tracking method in a clutter environment. Background Technology

[0002] Radar multi-target tracking tasks mainly involve data association and tracking filtering problems. The former refers to pairing radar measurement data at a certain moment with measurement data or tracks at other moments, while the latter refers to how to accurately estimate the motion state of the current target based on the current radar measurement data and previous measurement data. Traditional data association algorithms mainly include the Nearest Neighbor Algorithm (NN) based on distance metrics (such as Euclidean distance), which uses the nearest neighbor as the matching object. However, this algorithm performs poorly when targets are densely packed. The Probabilistic Data Association Algorithm (PDA) calculates the association probability between a measurement at a certain moment and a known target based on a Bayesian framework, updating and correcting the target state. This method has good results in handling single-target data association problems in clutter backgrounds. However, when the observation space is large and the number of targets is large, its computational complexity is high, and the effect is not ideal. The Joint Probabilistic Data Association (JPDA) algorithm is suitable for multi-target tracking in dense environments. While the underlying principle is similar, the association probability in JPDA is calculated from all measurements of all targets. The number of joint events increases exponentially with the number of targets, leading to an explosive increase in computational complexity. Therefore, this algorithm is unsuitable for practical engineering applications. All of the aforementioned radar multi-target data association algorithms rely on prior information such as target motion models, clutter density, filter covariance, and gate thresholds. However, in practical applications, it is impossible to accurately obtain this prior information beforehand.

[0003] Traditional tracking filtering algorithms are mainly divided into linear filtering and nonlinear filtering algorithms, such as the Kalman Filter Algorithm (KF) and the Extended Kalman Filter Algorithm (EKF), each with its own applicable scope. The quality of the filtering result is related to factors such as the statistical characteristics of process noise and measurement noise, and the initial conditions of the state. In reality, these quantities are unknown. If the assumed quantities used in filtering deviate significantly from the actual situation, it will lead to a completely incorrect estimation of the state. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method for generating a multi-target tracking model and a multi-target tracking method in cluttered environments. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] In a first aspect, embodiments of the present invention provide a method for generating a multi-target tracking model in a cluttered environment, comprising:

[0006] In a cluttered environment, the measured values ​​and true values ​​of a multi-target scene at each moment are obtained, and training data for the multi-target scene is obtained based on the measured values ​​and the true values; wherein, the measured values ​​include the measured data of the target position and the measured data of the clutter position, and the true values ​​are the true data of the target position;

[0007] A radar multi-target tracking network based on a cross-attention mechanism is constructed, comprising: a preprocessing module, a cross-attention mechanism layer, a forward propagation network, a residual structure, and a GRU network. For the current time t, the preprocessing module is used to obtain the parameters required by the cross-attention mechanism layer based on the estimated target state Outputs_t-1 at time t-1 and the measurement value Z_t at time t. Here, Outputs_t-1 is all zeros when t is the initial time, and Outputs_t-1 is the output of the radar multi-target tracking network at time t-1 when t is any time after the initial time. The cross-attention mechanism layer is used to obtain the data association probability between Outputs_t-1 and Z_t based on the parameters, thereby obtaining the target measurement information X at time t after clutter removal. atten The forward propagation network is used for X atten By performing dimensionality upscaling and nonlinear mapping, deep-level feature data X is obtained. FeedForword The residual structure is used to convert X atten and X FeedForword The sum of the target measurement information X is obtained by adding them together. t The GRU network is used for X t The output is the estimated state of the target at time t, Outputs_t, after filtering, and is used as the estimated position of the target at time t.

[0008] The radar multi-target tracking network is trained using the training data of the multi-target scenario to obtain a trained multi-target tracking model; wherein, the multi-target tracking model is used to track the positions of multiple targets.

[0009] In one embodiment of the present invention, acquiring the measured value and the actual value at each moment corresponding to a multi-target scene in a cluttered environment includes:

[0010] By setting simulation parameters, the measured values ​​and true values ​​at each moment are obtained through simulation. The simulation parameters include: the number of sampling points N, the sampling interval T, the number of targets, the initial state of the targets, the motion mode of the targets, Gaussian white noise, and clutter points following a Poisson distribution. The measured values ​​contain noise, while the true values ​​do not contain noise.

[0011] In one embodiment of the present invention, training data for a multi-target scene is obtained based on the measured values ​​and the true values, including:

[0012] For measurement values ​​where the sum of the number of clutter points and the number of target points does not reach the preset maximum number of measurements, fill in 0 values ​​to complete the dimension processing;

[0013] Normalize all the measured values ​​and true values ​​obtained so far to obtain normalized measured values ​​and true values;

[0014] The normalized measurement value and the true value at each time step are used as a set of data. The training data for the multi-objective scene is obtained based on the sets of data obtained at all times.

[0015] In one embodiment of the present invention, the process by which the preprocessing module obtains the parameters required by the cross-attention mechanism layer based on the estimated state Outputs_t-1 of the target at time t-1 and the measurement value Z_t at time t includes:

[0016] The estimated state Outputs_t-1 of the target is increased in dimensionality using the linear layer Linear_q in the preprocessing module to obtain the query matrix Q, where Q = Linear_q(Outputs_t-1), and Linear_q represents linear transformation processing.

[0017] The measurement value Z_t is increased in dimension using the linear layer Linear_k in the preprocessing module to obtain the key matrix K, K = Linear_k(Z_t), where Linear_k represents linear transformation processing;

[0018] The measured value Z_t is used as the value matrix V, V = Z_t;

[0019] The query matrix Q, the key matrix K, and the value matrix V are used as the parameters required for the cross-attention mechanism layer.

[0020] In one embodiment of the present invention, the target measurement information X atten It is determined using the first formula, which is: softmax represents the normalized exponential function, mask represents masking, Cross_Attention represents the cross-attention mechanism, ⊙ represents the Hadamard product, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k This represents the dimension of the key matrix K.

[0021] In one embodiment of the present invention, deep feature data X FeedForword It is determined using the second formula, which is X. FeedForword =Linear_2(ReLu(Linear_1(X) atten )));where Linear_1 represents the first layer of linear transformation processing, ReLU represents the activation function, and Linear_2 represents the second layer of linear transformation processing.

[0022] In one embodiment of the present invention, the GRU network targets X. t The filtered output is the estimated state of the target at time t, Outputs_t, which serves as the estimated position of the target at time t, including:

[0023] The GRU network is based on the sum of the target measurement-related information X. t Get updated gate z t , z t =σ(W z ·[Outputs_t-1, X t ]);

[0024] The GRU network is based on the sum of the target measurement-related information X. t Get the reset gate r t r t =σ(W r ·[Outputs_t-1, X t ]);

[0025] The GRU network is based on the sum of the target measurement-related information X. t and the reset gate r t Obtain candidate hidden state

[0026] The GRU network considers the candidate hidden state. Perform an update to obtain the hidden state at each time step, and update the hidden state h′ at the last time step. t Perform fully connected layer processing to obtain the estimated state of the target at time t, Outputs_t, which is taken as the target's position. Outputs_t = (1-z) t )⊙Outputs_t-1+z t ⊙h′ tAmong them, z t Indicates the update gate, r t This indicates that the door is being reset. W represents the candidate hidden state. z W represents the weight matrix of the updated gate. r This represents the weight matrix of the reset gate, σ represents the Sigmoid function, and W... h Let represent the weight matrix of the candidate hidden state, tanh represent the hyperbolic tangent function, and ⊙ represent the Hadamard product.

[0027] In one embodiment of the present invention, the radar multi-target tracking network is trained using training data from the multi-target scenario, and this is achieved using pre-set network parameters; wherein, the network parameters include:

[0028] The number of training cycles (epochs), the batch size (batch_size), and the dimension d of the key matrix K in the parameters. k The dimensions d_ff of the linear layer in the forward propagation network and the dimensions hidden_dim of the hidden layer in the GRU network.

[0029] Secondly, embodiments of the present invention provide a multi-target tracking method in a cluttered environment, including:

[0030] Utilizing radar to acquire measurement values ​​in multi-target scenarios;

[0031] The measured values ​​are input into the multi-target tracking model to obtain the estimated positions of all targets; the multi-target tracking model is obtained based on the steps of the multi-target tracking model generation method under clutter environment described in the first aspect.

[0032] The beneficial effects of this invention are:

[0033] In the solution provided by this invention, the multi-target tracking model can learn the data association probability between the estimated state at the previous moment and the measurement value at the current moment through the cross-attention mechanism layer without any prior information, thereby obtaining the target measurement information at the current moment after removing clutter; and it can use the GRU network to filter the sum of target measurement related information obtained through residual connections to obtain the estimated value of the target position at the current moment, thus realizing the tracking of multiple targets; compared with the prior art, when using the trained multi-target tracking model to perform multi-target tracking tasks, no parameter values ​​need to be adjusted, and the tracking effect is better. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for generating a multi-target tracking model in a cluttered environment, as provided in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the structure of a multi-target tracking model in a cluttered environment provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of a GRU network for a multi-target tracking model in a cluttered environment, provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram of the actual trajectory of a target in a cluttered environment, provided by an embodiment of the present invention.

[0038] Figure 5 The image shows the tracking results of the existing JPDA multi-target tracking algorithm.

[0039] Figure 6 The image shows the tracking results of a multi-target tracking model in a cluttered environment, as provided in an embodiment of the present invention.

[0040] Figure 7 The image shows a comparison of the OSPA distance between the tracking results of a multi-target tracking model in a cluttered environment provided in this embodiment of the invention and the tracking results of the existing JPDA multi-target tracking technology.

[0041] Figure 8 This is a flowchart illustrating a multi-target tracking method in a cluttered environment, as provided in an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Traditional radar multi-target data association algorithms rely on prior information such as target motion models, clutter density, filter covariance, and gate thresholds. The quality of filtering results in traditional tracking filtering algorithms is related to factors such as the statistical characteristics of process noise and measurement noise, and initial state conditions. However, this prior information is often unavailable in real-world radar multi-target tracking tasks. Therefore, to address these issues, this invention proposes a multi-target tracking model generation method and a multi-target tracking method for clutter environments. In clutter environments, this method requires no prior information and achieves better tracking robustness than traditional methods. Furthermore, the proposed radar multi-target tracking network can solve the association and matching problem for multiple targets and multiple measurements in any situation where target tracks intersect.

[0044] Example 1

[0045] This invention provides a method for generating a multi-target tracking model in a cluttered environment, such as... Figure 1 As shown, it may include:

[0046] S1, acquire the measurement and true values ​​of the multi-target scene at each time step in a cluttered environment, and obtain the training data of the multi-target scene based on the measurement and true values; wherein, the measurement values ​​include the measurement data of the target position and the measurement data of the clutter position, and the true values ​​are the true data of the target position;

[0047] S2. Construct a radar multi-target tracking network based on a cross-attention mechanism, including: a preprocessing module, a cross-attention mechanism layer, a forward propagation network, a residual structure, and a GRU network. For the current time t, the preprocessing module is used to obtain the parameters required by the cross-attention mechanism layer based on the estimated target state Outputs_t-1 at time t-1 and the measurement value Z_t at time t. Here, when t is the initial time, Outputs_t-1 is all 0s; when t is any time after the initial time, Outputs_t-1 is the output of the radar multi-target tracking network at time t-1. The cross-attention mechanism layer is used to obtain the data association probability between Outputs_t-1 and Z_t based on the parameters, thereby obtaining the target measurement information X at time t after clutter removal. atten Forward propagation networks are used to analyze X. atten By performing dimensionality upscaling and nonlinear mapping, deep-level feature data X is obtained. FeedForword The residual structure is used to convert X atten and X FeedForword The sum of the target measurement information X is obtained by adding them together. t GRU networks are used to analyze X. t The output is the estimated state of the target at time t, Outputs_t, after filtering, and is used as the estimated position of the target at time t.

[0048] S3. The radar multi-target tracking network is trained using training data from multi-target scenarios to obtain a trained multi-target tracking model; the multi-target tracking model is used to track the positions of multiple targets.

[0049] The multi-target tracking model provided in this embodiment of the invention, compared with existing classical algorithms, can learn the data association probability between the estimated state of the previous moment and the measurement value of the current moment through the cross-attention mechanism layer without any prior information, thereby obtaining the target measurement information of the current moment after removing clutter; and can use the GRU network to filter the sum of target measurement related information obtained through residual connection to obtain the position of the target at the current moment, thus realizing the tracking of multiple targets.

[0050] For ease of understanding, the following describes each step of the method for generating a multi-target tracking model in a cluttered environment proposed in the embodiments of the present invention.

[0051] For S1, it can include:

[0052] S1A acquires the measured and true values ​​at each moment in a multi-target scenario under cluttered conditions; the measured values ​​include the measured data of the target location and the measured data of the clutter location, and the true values ​​are the true data of the target location.

[0053] Specifically, for S1A, it can include:

[0054] By setting simulation parameters, the measured values ​​and true values ​​at each moment are obtained through simulation. The simulation parameters include: the number of sampling points N, the sampling interval T, the number of targets, the initial state of the targets, the motion mode of the targets, Gaussian white noise, and clutter points that follow a Poisson distribution. The measured values ​​contain noise, while the true values ​​do not.

[0055] S1A can be understood as the initial data generation step, which generates multiple multi-target scenarios under clutter interference through simulation, thereby obtaining the measured values ​​and true values ​​at each moment. Here, the number of sampling points N and the sampling interval T represent the frequency of data acquisition during simulation; the number of targets represents the number of targets in the generated multi-target scenario; the initial state of the target represents the initial position of the target; and the target's motion mode represents the target's movement mode, which can be set to a random mode. Clutter points following a Poisson distribution are uniformly distributed within the radar scanning area. It is understood that the method of obtaining target measured values ​​and true values ​​through simulation is only one feasible approach and should not be construed as a limitation on the data acquisition method of this embodiment. The data acquisition method can also be to scan targets with known true positions using radar, using the actual scanned results as measured values ​​and the true positions as true values. When simulating and generating multiple multi-target scenarios, each scenario contains multiple targets, each target adopts a different motion mode, and the generated trajectory is used as the target's true trajectory. Each trajectory contains the target's position information at multiple times, which is used as the target's true value. Gaussian noise is added to the true trajectory to simulate the target's measurement value. Clutter points, whose number follows a Poisson distribution and are uniformly distributed in the multi-target motion scenario, are added to simulate the interference of the external environment. The clutter point positions and target measurement values ​​contained at each time are used as the measurement value.

[0056] S1B is training data for multi-target scenarios based on measurements and real values.

[0057] Specifically, for S1B, it can include:

[0058] 1) Fill the measurement values ​​where the sum of the number of clutter points and the number of target points does not reach the preset maximum number of measurements with 0 values ​​to complete the dimension processing.

[0059] For each measurement value where the sum of the number of clutter points and target points does not reach the preset maximum number of measurements, the measurement value is filled with 0 values ​​so that the sum of the number of measurement values ​​is equal to the preset maximum number of measurements, thereby completing the dimension processing and obtaining the processed measurement value.

[0060] Considering the clutter interference in real-world scenarios, the number of clutter points varies at each moment. Therefore, it is necessary to process the dimensions of the target measurement values ​​with clutter to ensure that the dimensions of the processed measurement values ​​are the same.

[0061] 2) Normalize all the measured values ​​and true values ​​obtained so far to obtain normalized measured values ​​and true values;

[0062] Optionally, min-max normalization can be used to normalize the processed measurement values ​​and the true values; the formula for min-max normalization is as follows:

[0063] Where o′ represents the processed data, o represents the data that needs to be processed, min represents the minimum value calculated for all the data that needs to be processed, and max represents the maximum value calculated for all the data that needs to be processed.

[0064] 3) The normalized measurement value and the true value at each time step are used as a set of data. Based on the sets of data obtained at all times, training data for the multi-target scenario is obtained.

[0065] Data obtained at all time points can be used as training data for multi-target scenarios. Optionally, the data can be divided into training, validation, and test sets according to a certain ratio for training, validation, and testing of the subsequently constructed radar multi-target tracking network based on the cross-attention mechanism. This ratio can be selected as needed, and can be 7:2:1.

[0066] For S2, a radar multi-target tracking network based on a cross-attention mechanism is constructed, which may include: a preprocessing module, a cross-attention mechanism layer, a forward propagation network, a residual structure, and a GRU network; for the structure of the radar multi-target tracking network based on the cross-attention mechanism, please refer to [link to relevant documentation]. Figure 2 understand.

[0067] The following describes the working process of each module in a radar multi-target tracking network, which may include:

[0068] S2A, for the current time t, the preprocessing module obtains the parameters required by the cross-attention mechanism layer based on the estimated state Outputs_t-1 of the target at time t-1 and the measurement value Z_t at time t;

[0069] For the processing of the radar multi-target tracking network at time t, its input data are the estimated state of the target at time t-1, Outputs_t-1, and the measurement value Z_t at time t; where, when t is the initial time, Outputs_t-1 is all 0, and when t is any time after the initial time, Outputs_t-1 is output by the radar multi-target tracking network; the measurement value Z_t at time t can be obtained from step S1.

[0070] Specifically, S2A may include:

[0071] The estimated state Outputs_t-1 of the target is increased in dimension using the linear layer Linear_q in the preprocessing module to obtain the query matrix Q, Q = Linear_q(Outputs_t-1), where Linear_q represents the linear transformation process, that is, the linear transformation process of the linear layer Linear_q.

[0072] The measurement value Z_t is increased in dimension using the linear layer Linear_k in the preprocessing module to obtain the key matrix K, K = Linear_k(Z_t), where Linear_k represents the linear transformation process, i.e. the linear transformation process of the linear layer Linear_k;

[0073] The measured value Z_t is used as the value matrix V, V = Z_t;

[0074] The query matrix Q, key matrix K, and value matrix V are used as the parameters required for the cross-attention mechanism layer.

[0075] Among them, the query matrix Q, the key matrix K, and the value matrix V are variable parameters in the cross-attention mechanism, which are specifically related to the input of the radar multi-target tracking network. The query matrix Q and the key matrix K are used to calculate the subsequent attention weights, and the value matrix V directly represents the measurement value Z_t.

[0076] The inputs to the radar multi-target tracking network are the estimated target state Outputs_t-1 at time t-1 and the measurement value Z_t at time t. Outputs_t-1 and Z_t are then subjected to dimensionality increases through linear layers Linear_q and Linear_k, respectively, to obtain the query matrix Q and key matrix K in the cross-attention mechanism layer. Finally, Z_t is directly used as the value matrix V to complete the data preprocessing operation.

[0077] In S2B, the cross-attention mechanism layer obtains the data association probability between Outputs_t-1 and Z_t based on the parameters, thereby obtaining the target measurement information X at time t after removing clutter. atten ;

[0078] Compared to self-attention mechanisms, which only focus on the relationships between their own sequences, cross-attention mechanisms allow a sequence to focus on another sequence and use a dot product to calculate the attention weights. This precisely corresponds to the essence of data association in multi-target tracking. First, the measurement value Z_t-1 at time t-1 is processed through data association and filtering to obtain the estimated state Outputs_t-1 at time t-1. Then, it is correlated with the measurement value Z_t at time t to obtain the data association probability. The estimated state Outputs_t-1 at time t-1 can be represented as one sequence, and the measurement value Z_t at time t is another sequence. The data association probability corresponds to the attention weights, and the dot product between the data association probability and the measurement value Z_t at time t yields the target measurement information X at time t after clutter removal. atten The target measurement information actually excludes clutter and only contains the target location. The data is expanded into high-dimensional data through the cross-attention mechanism layer, making it easier to distinguish between target, clutter, and noise information, thereby increasing the target weight and decreasing the weight of clutter and noise, ultimately extracting the optimal data association probability.

[0079] Data association probability reflects the degree of correlation between the radar measurement value at the current moment and the existing target tracking results. In this embodiment of the invention, the greater the association probability value between a certain data point in the current measurement value and the estimated position of a target at the previous moment, the more likely that data point is the current measurement value of the target.

[0080] Specifically, target measurement information X atten It is determined using the first formula, which is: softmax represents the normalized exponential function, mask represents masking, Cross_Attention represents the cross-attention mechanism, ⊙ represents the Hadamard product, and d k Q represents the dimension of the key matrix K, Q represents the query matrix, K represents the key matrix, and V represents the value matrix.

[0081] The specific operation of `mask` involves creating a mask matrix, where the mask value is 1 for positions filled with 0 in the corresponding measurement value Z_t, and 0 for the remaining positions. This mask matrix is ​​then converted to Boolean values, meaning positions with a mask value of 1 are True, and positions with a mask value of 0 are False. The value corresponding to the True position in the mask matrix is ​​replaced with negative infinity. After the negative infinity is processed by the subsequent normalized exponential function, the output result is 0, thus eliminating the influence of the padding value.

[0082] S2C, forward propagation network for X atten By performing dimensionality upscaling and nonlinear mapping, deep-level feature data X is obtained. FeedForword ;

[0083] Specifically, deep feature data X FeedForword It is determined using the second formula, which is X. FeedForword =Linear_2(ReLu(Linear_1(X) atten )));where Linear_1 represents the first layer of linear transformation processing, ReLU represents the activation function, and Linear_2 represents the second layer of linear transformation processing.

[0084] S2D, residual structure will X atten and X FeedForword The sum of the target measurement information X is obtained by adding them together. t ;

[0085] Specifically, the residual structure in Figure 2 China and Israel X represents the sum of target measurement-related information. t It is determined using the third formula, which is X. t =X atten +X FeedForword .

[0086] The introduced residual structure is used to prevent problems such as gradient explosion, vanishing gradients, and network degradation; the target measurement information X is then used to... atten and deep feature data X FeedForword Add them together to obtain the sum of target measurement information X. t This serves as the input to the subsequent GRU network. Specifically, the target measurement information X... atten Without processing through a forward propagation network, deep feature data X FeedForword This is the output obtained after processing by the forward propagation network.

[0087] In the S2A-S2D process described above, the data obtained after the dot product of the data association probability and the measurement value Z_t at time t only shows the estimated state of the target at time t-1 (Outputs_t-1), thus only providing short-term information about the target and making it impossible to know the target's motion characteristics. However, the filtering and tracking problem is closely related to the target's motion characteristics. Therefore, to enable the radar multi-target tracking network to learn the target's long-term motion information, a network suitable for solving time series prediction needs to be introduced. Considering both tracking accuracy and timeliness, this embodiment of the invention can choose to introduce a GRU network for subsequent processing.

[0088] S2E, GRU network to Xt The output, Outputs_t, is the estimated state of the target at time t after filtering, and is used as the target's position.

[0089] Specifically, S2E can include:

[0090] The GRU network calculates the sum of target measurement information X. t Get updated gate z t , z t =σ(W z ·[Outputs_t-1, X t ]);

[0091] The GRU network calculates the sum of target measurement information X. t Get the reset gate r t r t =σ(W r ·[Outputs_t-1, X t ]);

[0092] The GRU network calculates the sum of target measurement information X. t and reset door r t Obtain candidate hidden state

[0093] GRU network for candidate hidden states Perform an update to obtain the hidden state at each time step, and update the hidden state h′ at the last time step. t Perform fully connected layer processing to obtain the estimated state Outputs_t at time t, which is taken as the target's position. Outputs_t = (1-z) t )⊙Outputs_t-1+z t ⊙h′ t Among them, z t Indicates the update gate, r t This indicates that the door is being reset. W represents the candidate hidden state. z W represents the weight matrix of the updated gate. r This represents the weight matrix of the reset gate, σ represents the Sigmoid function, and W... h Let represent the weight matrix of the candidate hidden state, tanh represent the hyperbolic tangent function, and ⊙ represent the Hadamard product.

[0094] Please refer to the structural diagram of the GRU network. Figure 3 .from Figure 3 As can be seen from this, by updating gate z t Reset door r t and candidate hidden state The hidden state at each time step is obtained by processing the hidden state and updating the candidate hidden state. The hidden state at the last time step is taken and processed through a fully connected layer to obtain the estimated state Outputs_t at time t, which is used as the position of the target.

[0095] The mechanism of GRU network is to use a gating mechanism to capture the dependencies in the sequence data, reset the gate to capture short-term dependencies, and update the gate to capture long-term dependencies (usually the data closer to time t has stronger correlation). The measurement data before time t are all contained between the hidden states, and the network output corresponds to the estimated state at time t. This is consistent with the filtering concept in target tracking and is suitable for filtering tracking tasks.

[0096] To better understand the various modules of the radar multi-target tracking network proposed in this embodiment of the invention, the following description is provided in conjunction with specific training steps.

[0097] For S3, the radar multi-target tracking network is trained using training data from multi-target scenarios to obtain a trained multi-target tracking model; the multi-target tracking model is used to track the positions of multiple targets.

[0098] The radar multi-target tracking network is trained using training data from multi-target scenarios and implemented using pre-set network parameters.

[0099] Specifically, for S3, the following steps may be included:

[0100] S3A sets the network parameters and inputs the estimated target state Outputs_t-1 at time t-1 and the measurement value Z_t at time t in the training data of the multi-target scenario into the radar multi-target tracking network.

[0101] Network parameters may include:

[0102] The number of training cycles (epochs), the batch size (batch_size), and the dimension d of the key matrix K in the parameters. k The dimensions of the linear layers in the forward propagation network are d_ff and the dimensions of the hidden layers in the GRU network are hidden_dim. The training samples consist of the input to the radar multi-target tracking network and the corresponding label values, with the target's true value used as the label value Y_t.

[0103] S3B, the preprocessing module preprocesses the estimated state Outputs_t-1 of the target at time t-1 and the measurement value Z_t at time t to obtain the parameters required by the cross-attention mechanism layer:

[0104] Q = Linear_q(Outputs_t-1);

[0105] K = Linear_k(Z_t);

[0106] V = Z_t; where Q represents the query matrix, K represents the key matrix, V represents the value matrix, Linear_q represents linear transformation, and Linear_k represents linear transformation.

[0107] S3C, the cross-attention mechanism layer processes the parameters it receives to obtain target measurement information X. atten :

[0108] softmax represents the normalized exponential function, mask represents masking, Cross_Attention represents the cross-attention mechanism, ⊙ represents the Hadamard product, and d k Q represents the dimension of the key matrix K, Q represents the query matrix, K represents the key matrix, and V represents the value matrix.

[0109] S3D, the forward propagation network for target measurement information X atten The data is processed to obtain deep feature data X. FeedForword :

[0110] X FeedForword =Linear_2(ReLu(Linear_1(X) atten )));where Linear_1 represents the first layer of linear transformation processing, ReLU represents the activation function, and Linear_2 represents the second layer of linear transformation processing.

[0111] S3E, residual structure to target measurement information X atten and deep feature data X FeedForword The residual connection processing yields the sum of target measurement related information X. t .

[0112] S3F, summing the target measurement information X t Input to the GRU network to obtain update gate z t Reset door r t Candidate hidden state And update the hidden state to get the target's position Outputs_t:

[0113] z t =σ(W z ·[Outputs_t-1, X t ]);

[0114] r t =σ(W r ·[Outputs_t-1, X t ]);

[0115]

[0116] Outputs_t = (1-z) t )⊙Outputs_t-1+z t ⊙h′ t Among them, z t Indicates the update gate, r t This indicates that the door is being reset. W represents the candidate hidden state. z W represents the weight matrix of the updated gate. r This represents the weight matrix of the reset gate, σ represents the Sigmoid function, and W... h Let represent the weight matrix of the candidate hidden state, tanh represent the hyperbolic tangent function, and ⊙ represent the Hadamard product.

[0117] S3F optimizes the radar multi-target tracking network using a loss function, resulting in a trained multi-target tracking model.

[0118] Specifically, the root mean square error between the target position Outputs_t at time t and the label value Y_t is obtained through the fourth formula. In the training data of the multi-target scenario, the loss function of the training set is used for batch gradient descent, while the loss function of the validation set is only used to evaluate the model's generalization ability and does not participate in gradient descent. The fourth formula is as follows:

[0119] Where M represents the duration of the target trajectory.

[0120] The radar multi-target tracking network performs backpropagation and is continuously trained until the loss function decreases until the radar multi-target tracking network converges, thus completing the training process and obtaining the trained multi-target tracking model.

[0121] After obtaining the trained multi-target tracking model, the tracking performance of the multi-target tracking model can be tested using the test set in the training data of the multi-target scene. If the tracking performance meets expectations, the multi-target tracking model can be directly used to track targets without needing to adjust parameters. If the tracking performance does not meet expectations, the target tracking model can be trained again until the tracking performance meets expectations.

[0122] To more clearly demonstrate the superior performance of the radar multi-target tracking network proposed in this embodiment of the invention, a simulation comparison is performed below between the existing JPDA multi-target tracking algorithm and the radar multi-target tracking network proposed in this embodiment of the invention.

[0123] The hardware configuration for the experiment was as follows: CPU: AMD Ryzen 75800X 8-Core Processor; GPU: GeForce RTX 3070Ti; RAM: 16GB; CUDA 11.2 was used to accelerate training by utilizing the GPU.

[0124] Experimental data generation: With sampling points N=64, sampling interval T=1, and target number 4, 500 multi-target scenarios were simulated. Each scenario contained 4 targets, with random initial positions and motion patterns for each target. Target measurements were affected by noise, so Gaussian white noise with a mean of 0 and a variance of 10 was added. The number of clutter points in each scenario followed a Poisson distribution with a mean of λ, and a uniform distribution within the radar scanning area. For a target trajectory diagram of one of the multi-target scenarios, please refer to [link to relevant documentation]. Figure 4 Four targets move at a constant speed in a straight line within the region with different initial positions. The solid line represents the actual trajectory of the target, and the solid dots represent measurement points, including the target's measurement points and clutter points. Target 4 intersects with targets 1 and 3, and target 2 intersects with the other targets.

[0125] Network parameter settings: During network training, the SGD optimizer is used for parameter updates, with an initial learning rate of 0.001, 1000 training epochs, a batch size of 64 samples per training iteration, and a dimension d of the key matrix K. k The dimension d_ff of the linear layer in the forward propagation network is 1024, and the dimension hidden_dim of the hidden layer in the GRU network is 64.

[0126] Evaluation Metrics: This invention uses Optimal Sub-Patten Assignment (OSPA) to evaluate the multi-target tracking performance of the algorithm. OSPA can quantitatively assess the difference between the actual trajectory and the estimated trajectory in a multi-target tracking system. Its formula is defined as follows:

[0127] Where Outputs represents the output of the JPDA multi-target tracking algorithm or the radar multi-target tracking network proposed in this embodiment of the invention, Y represents the true value of the target, n represents the estimated number of targets, m represents the true number of targets, and π represents the true number of targets. n d represents the number of all permutations and combinations of taking m elements from Y. c (Outputs i Y π(i) ) = min(C, ||Y i -Outputs i ||), |||| denotes the 2-norm. Let p represent the group of estimated states and actual states of all targets with the smallest distance, where p (1≤p≤∞) represents the distance sensitivity parameter and C (C>0) represents the association sensitivity parameter.

[0128] Experimental Results: To demonstrate the effectiveness of the radar multi-target tracking network, it was compared with the existing JPDA multi-target tracking algorithm in a simulation experiment. Figure 5 The tracking results of the JPDA multi-target tracking algorithm are shown in the figure with a detection probability of 0.99 and a gate threshold of 30. It can be seen that the JPDA multi-target tracking algorithm cannot clearly separate the tracks of target 1 and target 2 at the intersection and in the part where target 2 and target 3 are close to each other. Figure 6 The image shows the tracking results of the radar multi-target tracking network provided in this embodiment of the invention. It can be seen that when target tracks intersect, the radar multi-target tracking network can accurately separate them. The OSPA distances of the two methods are calculated and compared. For a comparison of the OSPA distances between the tracking results of the multi-target tracking model proposed in this embodiment and the existing JPDA multi-target tracking technology, please refer to the image. Figure 7 It can be seen that the OSPA distance of the JPDA multi-target tracking algorithm is much greater than that of the radar multi-target tracking network when targets are close together and intersect. The OSPA distance of the radar multi-target tracking network is generally smaller, proving that the effectiveness and tracking performance of the radar multi-target tracking network are significantly better than the existing JPDA multi-target tracking algorithm. The data association probability obtained by the radar multi-target tracking network is sufficiently accurate. Especially in the case of intersecting multi-target trajectories, the tracking results of the radar multi-target tracking network proposed in this embodiment are far superior to the existing JPDA multi-target tracking algorithm.

[0129] In the solution provided by this invention, the multi-target tracking model can learn the data association probability between the estimated state at the previous moment and the measurement value at the current moment through the cross-attention mechanism layer without any prior information, thereby obtaining the target measurement information at the current moment after removing clutter; and it can use the GRU network to filter the sum of target measurement related information obtained through residual connections to obtain the position of the target at the current moment, thus realizing the tracking of multiple targets; compared with the prior art, when using the trained multi-target tracking model for multi-target tracking tasks, since the multi-target tracking model has been pre-trained, no parameter values ​​need to be adjusted in actual use, and the tracking effect is better.

[0130] Example 2

[0131] This invention provides a multi-target tracking method in cluttered environments, such as... Figure 8 As shown, it may include:

[0132] The S100 uses radar to acquire measurement values ​​in multi-target scenarios; the measurement values ​​include measurement data of target positions and measurement data of clutter positions.

[0133] S200, input the measurement values ​​into the multi-target tracking model to obtain the estimated values ​​of all target positions; the multi-target tracking model is obtained based on the steps of the multi-target tracking model generation method in clutter environment of Example 1.

[0134] For details on the target tracking model, please refer to Example 1 for further explanation; it will not be elaborated upon here.

[0135] The multi-target tracking method provided in this invention obtains the target positions in a multi-target scene as measurement values ​​through radar scanning. When processing the measurement values ​​using the multi-target tracking model provided in Embodiment 1, it can learn the data association probability between the estimated state at the previous moment and the measurement value at the current moment through a cross-attention mechanism layer without any prior information, thereby obtaining the target measurement information at the current moment after removing clutter. Furthermore, it can use a GRU network to filter the sum of target measurement-related information obtained through residual connections to obtain the target position at the current moment, thus achieving multi-target tracking. Since the multi-target tracking model has been pre-trained, no parameter values ​​need to be adjusted when applying the multi-target tracking method to it, and it can be used directly.

[0136] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0137] The above description is merely a preferred embodiment of the present invention and is 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 are included within the scope of protection of the present invention.

Claims

1. A method for generating a multi-target tracking model in a cluttered environment, characterized in that, include: In a cluttered environment, the measured values ​​and true values ​​of a multi-target scene at each moment are obtained, and training data for the multi-target scene is obtained based on the measured values ​​and the true values; wherein, the measured values ​​include the measured data of the target position and the measured data of the clutter position, and the true values ​​are the true data of the target position; A radar multi-target tracking network based on a cross-attention mechanism is constructed, comprising: a preprocessing module, a cross-attention mechanism layer, a forward propagation network, a residual structure, and a GRU network. For the current time t, the preprocessing module is used to obtain the parameters required by the cross-attention mechanism layer based on the estimated target state Outputs_t-1 at time t-1 and the measurement value Z_t at time t. Here, Outputs_t-1 is all zeros when t is the initial time, and Outputs_t-1 is the output of the radar multi-target tracking network at time t-1 when t is any time after the initial time. The cross-attention mechanism layer is used to obtain the data association probability between Outputs_t-1 and Z_t based on the parameters, thereby obtaining the target measurement information X at time t after clutter removal. atten The forward propagation network is used for X atten By performing dimensionality upscaling and nonlinear mapping, deep-level feature data X is obtained. FeedForword The residual structure is used to convert X atten and X FeedForword The sum of the target measurement information X is obtained by adding them together. t The GRU network is used for X t The output is the estimated state of the target at time t, Outputs_t, after filtering, and is used as the estimated position of the target at time t. The radar multi-target tracking network is trained using the training data of the multi-target scenario to obtain a trained multi-target tracking model; wherein, the multi-target tracking model is used to track the positions of multiple targets.

2. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The acquisition of the measured and actual values ​​at each moment in a multi-target scene under cluttered conditions includes: By setting simulation parameters, the measured values ​​and true values ​​at each moment are obtained through simulation. The simulation parameters include: the number of sampling points N, the sampling interval T, the number of targets, the initial state of the targets, the motion mode of the targets, Gaussian white noise, and clutter points following a Poisson distribution. The measured values ​​contain noise, while the true values ​​do not contain noise.

3. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, Training data for a multi-target scene is obtained based on the measured values ​​and the true values, including: For measurement values ​​where the sum of the number of clutter points and the number of target points does not reach the preset maximum number of measurements, fill in 0 values ​​to complete the dimension processing; Normalize all the measured values ​​and true values ​​obtained so far to obtain normalized measured values ​​and true values; The normalized measurement value and the true value at each time step are used as a set of data. The training data for the multi-objective scene is obtained based on the sets of data obtained at all times.

4. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The process by which the preprocessing module obtains the parameters required by the cross-attention mechanism layer based on the estimated state Outputs_t-1 of the target at time t-1 and the measurement value Z_t at time t includes: The estimated state Outputs_t-1 of the target is increased in dimensionality using the linear layer Linear_q in the preprocessing module to obtain the query matrix Q, where Q = Linear_q(Outputs_t-1), and Linear_q represents linear transformation processing. The measurement value Z_t is increased in dimension using the linear layer Linear_k in the preprocessing module to obtain the key matrix K, K = Linear_k(Z_t), where Linear_k represents linear transformation processing; The measured value Z_t is used as the value matrix V, V = Z_t; The query matrix Q, the key matrix K, and the value matrix V are used as the parameters required for the cross-attention mechanism layer.

5. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The target measurement information X atten It is determined using the first formula, which is: softmax represents the normalized exponential function, mask represents masking, Cross_Attention represents the cross-attention mechanism, ⊙ represents the Hadamard product, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k This represents the dimension of the key matrix K.

6. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The deep feature data X FeedForword It is determined using the second formula, which is X. FeedForword =Linear_2(ReLu(Linear_1(Xatten))); where Linear_1 represents the first layer of linear transformation processing, ReLu represents the activation function, and Linear_2 represents the second layer of linear transformation processing.

7. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The GRU network supports X t The filtered output is the estimated state of the target at time t, Outputs_t, which serves as the estimated position of the target at time t, including: The GRU network is based on the sum of the target measurement-related information X. t Get updated gate z t , z t =σ(W z ·[Outputs_t-1, X t ]); The GRU network is based on the sum of the target measurement-related information X. t Get the reset gate r t r t =σ(W r ·[Outputs_t-1, X t ]); The GRU network is based on the sum of the target measurement-related information X. t and the reset gate r t Obtain candidate hidden state The GRU network considers the candidate hidden state. Perform an update to obtain the hidden state at each time step, and update the hidden state h at the last time step. t A fully connected layer is applied to obtain the estimated state of the target at time t, Outputs_t, which is taken as the target's position. Outputs_t = (1-z) / t. t )⊙Outputs_t-1+z t ⊙h t ′;wherein, z t Indicates the update gate, r t This indicates that the door is being reset. W represents the candidate hidden state. z W represents the weight matrix of the updated gate. r This represents the weight matrix of the reset gate, σ represents the Sigmoid function, and W... h Let represent the weight matrix of the candidate hidden state, tanh represent the hyperbolic tangent function, and ⊙ represent the Hadamard product.

8. The method for generating a multi-target tracking model in a cluttered environment according to claim 1, characterized in that, The radar multi-target tracking network is trained using training data from the multi-target scenario, employing pre-set network parameters; wherein, the network parameters include: The number of training cycles (epochs), the batch size (batch_size), and the dimension d of the key matrix K in the parameters. k The dimensions d_ff of the linear layer in the forward propagation network and the dimensions hidden_dim of the hidden layer in the GRU network.

9. A multi-target tracking method in a cluttered environment, characterized in that, include: Utilizing radar to acquire measurement values ​​in multi-target scenarios; The measured values ​​are input into the multi-target tracking model to obtain estimated values ​​of the positions of all targets; The multi-target tracking model is obtained based on the steps of the multi-target tracking model generation method under clutter environment as described in any one of claims 1 to 8.

Images