Multi-target tracking method

Abstract

The invention discloses a multi-target tracking method. The method comprises the steps that the number of tracking targets, target state parameters, observation data and actual parameter weights are acquired; according to the number of the tracking targets and the target state parameters, constructing a target state function to obtain the target state function; performing state space construction according to the target state function to obtain a target state space; constructing a mutual relation matrix according to the target state space to obtain the mutual relation matrix; and according to the target state space and the correlation matrix, carrying out maximum expected iteration operation, and when it is determined that the number of iterations is greater than a preset value, outputting an estimated state. The method can improve the multi-target tracking precision.

Description

Technical Field

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

[0002] At present, multi-target tracking technology plays an important role in many fields such as autonomous driving, intelligent monitoring, and behavior recognition. Especially in the field of autonomous driving, MOT technology is crucial for detecting and tracking other vehicles and pedestrians, while in video surveillance and human-computer interaction, multi-target tracking technology mainly relies on image processing technology to achieve target recognition and tracking. This technology has broad application prospects. With the continuous development of advanced technologies such as deep learning, its application in various fields will become more extensive and in-depth.

[0003] In one existing technology, the Bayesian inference method is applied to multi-target tracking in order to improve the tracking accuracy. However, this method faces many challenges in practical applications. Due to the limitations of observation data, such as low resolution, insufficient light, and occlusion of the target, the Bayesian inference method is not ideal in multi-target tracking. These problems are characterized by the correlation between data, which increases the complexity of the multi-target tracking problem and leads to low multi-target tracking accuracy. Summary of the invention

[0004] The present invention provides a multi-target tracking method to improve the multi-target tracking accuracy.

[0005] In a first aspect, in order to solve the above technical problem, the present invention provides a multi-target tracking method, comprising:

[0006] Obtain the number of tracked targets, target state parameters, observation data and actual parameter weights;

[0007] According to the number of tracking targets and the target state parameters, a target state function is constructed to obtain a target state function;

[0008] According to the target state function, a state space is constructed to obtain a target state space;

[0009] According to the target state space, a mutual relationship matrix is ​​constructed to obtain a mutual relationship matrix;

[0010] According to the target state space and the mutual relationship matrix, a maximum expectation iteration operation is performed, and when it is determined that the number of iterations is greater than a preset value, an estimated state is output.

[0011] In an optional implementation, a target state function is constructed according to the number of tracked targets and the target state parameter to obtain the target state function, including:

[0012] Performing a normalization operation according to the number of tracked targets and the target state parameter to obtain a normalized target state;

[0013] According to the normalized target state and the preset time interval, time division is performed to obtain a target state function.

[0014] In an optional implementation, constructing a state space according to the target state function to obtain a target state space includes:

[0015] According to the target state function, a time alignment operation is performed to obtain an aligned target state function;

[0016] According to the alignment target state function, the states are integrated at the same time to obtain the target state space.

[0017] In an optional implementation, a mutual relationship matrix is ​​constructed according to the target state space to obtain a mutual relationship matrix, including:

[0018] Calculating the target state distance according to the target state space to obtain the target state distance;

[0019] According to the target state distance, a matrix is ​​constructed to obtain a mutual relationship matrix;

[0020] The formula for calculating the target state distance is as follows:

[0021] r kl =|f k -f l |

[0022] Among them, r kl represents the state distance between target k and target l, f k represents the state function of target k, f l Represents the state function of target l.

[0023] In an optional implementation, a maximum expectation iteration operation is performed according to the target state space and the mutual relationship matrix, and when it is determined that the number of iterations is greater than a preset value, an estimated state is output, including:

[0024] According to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data;

[0025] According to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space, a posterior probability maximization operation is performed to obtain the optimal model state parameters;

[0026] Performing a state estimation operation according to the target state function, the observation data and the optimal model state parameters to obtain an estimated state;

[0027] According to the estimated state, construct a mutual relationship matrix to obtain a second relationship matrix;

[0028] Perform matrix distance calculation according to the second relationship matrix and the mutual relationship matrix to obtain updated distance data;

[0029] When the updated distance data is less than a set threshold or the number of iterations is greater than a preset iteration threshold, the iteration ends.

[0030] In an optional implementation, performing Dirichlet distribution calculation according to the target state space to obtain estimated parameter weight data includes:

[0031] According to the target state space, a target frequency statistical operation is performed to obtain target occurrence frequency data;

[0032] Perform a weight calculation operation according to the target occurrence number data, the preset Dirichlet hyperparameters and the dimension of the target state function to obtain estimated parameter weight data;

[0033] The weight calculation formula is as follows:

[0034]

[0035] Among them, n k =[n k (1),n k (2),…n k (d)…,n k (D)] T represents the number of times target k appears in the state space, n k (d) represents the number of times the target k appears in the dth state space of the state space, represents the total number of occurrences of all targets, M is the total number of state spaces, D is the dimension of the target state function f, α is the hyperparameter of the Dirichlet distribution, and θ kt Represents the parameter weight of target k at time t.

[0036] In an optional implementation, performing a posterior probability maximization operation according to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space to obtain the optimal model state parameters includes:

[0037] Constructing a state posterior probability formula according to the estimated parameter weight data and the target state space;

[0038] According to the state posterior probability formula, a maximization solution operation is performed to obtain the optimal model state parameters;

[0039] The state posterior probability formula is as follows:

[0040]

[0041] Among them, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the t-th iteration, θ represents the model parameters, θ (t) represents the parameter estimate of the tth iteration, N is the number of samples, z n is the nth target state, p(z n |x n ,θ (t) ) is the posterior probability of the nth target state based on the model parameters at the tth iteration, p(x n ,z n |θ) is the complete data likelihood function, x n represents observation data;

[0042] The maximization solution formula is as follows:

[0043]

[0044] in, represents the optimal model state parameter obtained at the tth iteration, represents the selection of θ that maximizes Q, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the tth iteration.

[0045] In an optional implementation, performing a state estimation operation according to the target state function, the observation data and the optimal model state parameter to obtain an estimated state includes:

[0046] Calculating the posterior probability according to the target state function and the optimal model state parameters to obtain posterior probability data;

[0047] Performing an update operation according to the posterior probability data and the observation data to obtain an estimated state;

[0048] The update calculation formula is as follows:

[0049]

[0050] in, is the updated value of the state function of the kth target in the t+1th iteration, γ(z ik ) is the posterior probability that the i-th data point belongs to the k-th state, z ik represents the i-th data point belonging to the k-th state, x i is the observed data, and n is the total number of observed data.

[0051] In an optional implementation, performing matrix distance calculation according to the second relationship matrix and the mutual relationship matrix to obtain updated distance data includes:

[0052] Performing a subtraction operation on the second relationship matrix and the mutual relationship matrix to obtain a relationship difference matrix;

[0053] According to the relationship difference matrix, a negative number is replaced by a positive operation to obtain a relationship absolute difference matrix;

[0054] According to the relationship absolute difference matrix, elements are summed to obtain updated distance data.

[0055] In a second aspect, the present invention provides a multi-target tracking system, comprising:

[0056] Input module, used to obtain the number of tracked targets, target state parameters and actual parameter weights;

[0057] A state function construction module is used to construct a target state function according to the number of tracking targets and the target state parameters to obtain a target state function;

[0058] A state space construction module, used to construct the state space according to the target state function to obtain the target state space;

[0059] A mutual relationship construction module, used for constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix;

[0060] The output module is used to perform a maximum expectation iteration operation according to the target state space and the mutual relationship matrix, and output an estimated state when it is determined that the number of iterations is greater than a preset value.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] The present invention provides a multi-target tracking method, which is characterized by comprising: obtaining the number of tracked targets, target state parameters, observation data and actual parameter weights; constructing a target state function according to the number of tracked targets and the target state parameters to obtain a target state function; constructing a state space according to the target state function to obtain a target state space; constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix; performing a maximum expected iteration operation according to the target state space and the mutual relationship matrix, and outputting an estimated state when it is determined that the number of iterations is greater than a preset value.

[0063] The multi-target tracking method provided by the present invention realizes the accurate construction of the target state function by comprehensively considering the number of tracked targets, target state parameters, observation data and actual parameter weights. This method first constructs the target state function according to the number of tracked targets and the target state parameters, and defines a detailed state model for each target. Then, based on the target state function, the state space is further constructed to provide a comprehensive description of the state change of each target. In addition, the present invention quantifies the mutual influence between different targets by constructing a mutual relationship matrix, which is crucial in multi-target tracking. The introduction of the mutual relationship matrix enables the algorithm to more accurately simulate the dynamic relationship between targets, especially when the targets are close or blocked. Finally, through the maximum expectation iteration operation, that is, the EM algorithm, the present invention continuously optimizes the estimation of the target state. The iterative process of the EM algorithm allows the algorithm to update the model parameters according to the current best estimate in each iteration, thereby gradually improving the accuracy of the estimation. When the number of iterations reaches a preset threshold, the algorithm outputs the final estimated state, which reflects the most likely position and state of the target during the tracking process. Therefore, the present invention significantly improves the accuracy of multi-target tracking by accurately constructing the target state function, state space and mutual relationship matrix, combined with the iterative optimization of the EM algorithm. This approach can not only handle the mutual influence between targets, but also adapt to the dynamic changes of target states, thereby providing more accurate and reliable tracking results in complex tracking scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of a multi-target tracking method provided by an embodiment of the present invention;

[0065] Figure 2 It is a schematic diagram of the structure of a multi-target tracking system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Reference Figure 1 , an embodiment of the present invention provides a multi-target tracking method, comprising the following steps:

[0068] S11, obtaining the number of tracked targets, target state parameters, observation data and actual parameter weights;

[0069] S12, constructing a target state function according to the number of tracked targets and the target state parameter to obtain a target state function;

[0070] S13, constructing a state space according to the target state function to obtain a target state space;

[0071] S14, constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix;

[0072] S15, performing a maximum expectation iteration operation according to the target state space and the mutual relationship matrix, and outputting an estimated state when it is determined that the number of iterations is greater than a preset value.

[0073] In step S11, the number of tracked targets, target state parameters, observation data and actual parameter weights are obtained. It should be noted that in the multi-target tracking method, the number of tracked targets needs to be obtained first, which is to determine the number of targets to be tracked, which can be identified and counted in the video frame or image by the target detection algorithm. The target state parameter refers to the representation of each target in the state space, including position, speed, acceleration, etc. These parameters can be obtained by target detection and feature extraction. Exemplarily, the present invention uses a deep learning model to predict the bounding box and motion features of the target. Of course, other methods can also be selected according to different application scenarios and user needs, and the present invention does not limit this. Observation data refers to actual data captured from a sensor or camera, which includes the appearance features, motion trajectory, etc. of the target, and is obtained through image processing and computer vision technology. The actual parameter weight refers to the weight factor used to adjust the importance of different parameters in the model.

[0074] In step S12, a target state function is constructed according to the number of tracked targets and the target state parameter to obtain a target state function, including:

[0075] Performing a normalization operation according to the number of tracked targets and the target state parameter to obtain a normalized target state;

[0076] According to the normalized target state and the preset time interval, time division is performed to obtain a target state function.

[0077] It should be noted that, in step S12, first, a normalization operation is performed according to the number of tracking targets and the target state parameter. Normalization is to adjust the value of the target state parameter to a standard range. For example, the present invention selects a normalization range of -1 to 1. Of course, according to different application scenarios and user needs, other normalization ranges can also be selected, and the present invention does not limit this. The purpose of doing so is to eliminate the influence of different parameter dimensions and numerical ranges, so that the algorithm has consistent sensitivity to different state parameters. For example, the normalization operation of the present invention is to subtract the minimum value and divide it by the difference between the maximum value and the minimum value to achieve. Of course, according to different application scenarios and user needs, other statistical methods, such as Z score normalization, can also be used, and the present invention does not limit this. After obtaining the normalized target state, time division is performed according to the normalized target state and the preset time interval. Time division refers to dividing the continuous time axis into discrete time units, each time unit corresponding to a time interval. This step is to model the target state in the time dimension so that the algorithm can independently process the change of the target state in each time unit. The choice of time interval depends on the specific requirements of the tracking task and the sampling rate of the sensor data. Exemplarily, the sampling frequency adopted by the present invention is 1000 times per second, that is, 1000Hz. Of course, other sampling frequencies can also be used according to different application scenarios and user needs, and the present invention is not limited to this. On the basis of time division, the target state function is obtained. The target state function is a mathematical model that describes the change of the target state over time. It can predict the state of the target at a future time point based on the normalized state and time information of the target. Exemplarily, the state function adopted by the present invention is linear. Of course, according to different application scenarios and user needs, nonlinear state space models or state prediction models based on machine learning can also be used, and the present invention is not limited to this. The construction of the target state function involves the estimation of parameters, which define the dynamic characteristics of the state transition.

[0078] It is worth noting that after the target state function is constructed, it will be used in the subsequent tracking process. In the context of multi-target tracking, the target state function can be used to predict the future position of the target, evaluate the interaction between targets, and update the estimate of the target state based on the observed data. In this way, the target state function provides a dynamic and time-related framework for the multi-target tracking algorithm, allowing the algorithm to more accurately handle the motion and mutual relationship of the targets.

[0079] In step S13, a state space is constructed according to the target state function to obtain a target state space, including:

[0080] According to the target state function, a time alignment operation is performed to obtain an aligned target state function;

[0081] According to the alignment target state function, the states are integrated at the same time to obtain the target state space.

[0082] It should be noted that, in step S13, first, a time alignment operation is performed according to the target state function. Time alignment refers to adjusting the state functions of different targets to the same time reference. This is done to ensure that when comparing or integrating the states of different targets, all state information is based on the same time point. Exemplarily, the time alignment operation of the present invention selects an interpolation method to ensure that the state function of each target is defined at the same time point. Of course, according to different application scenarios and user needs, methods such as extrapolation or state prediction can also be used, and the present invention is not limited to this. After obtaining the aligned target state function, the next step is to perform simultaneous state integration according to the aligned target state function. Simultaneous state integration refers to merging the state information of all targets at the same time point into a unified state space. This integration process needs to process the state information of multiple targets, including parameters such as position, velocity, acceleration, etc., and combine them into a multidimensional vector or matrix, in which each column represents the state of a target at that time point. Based on the simultaneous state integration, the target state space is obtained. The target state space is a high-dimensional space, in which each dimension represents a state parameter, and each point in the space represents the state of a target at a specific time point. This space allows the algorithm to model and predict the state of the target at consecutive time points, and can be used to analyze the interaction and relative motion between targets. After the target state space is constructed, it will be used in the subsequent tracking process. The target state space provides a unified framework for multi-target tracking algorithms, allowing the algorithm to process the state information of all targets simultaneously at each time point. This helps to improve the accuracy and robustness of tracking, especially when there are interactions between targets or in complex environments. With the target state space, the algorithm can perform data association, target identification, and state estimation more efficiently, thereby achieving more accurate multi-target tracking.

[0083] In step S14, a mutual relationship matrix is ​​constructed according to the target state space to obtain a mutual relationship matrix, including:

[0084] Calculating the target state distance according to the target state space to obtain the target state distance;

[0085] According to the target state distance, a matrix is ​​constructed to obtain a mutual relationship matrix;

[0086] The formula for calculating the target state distance is as follows:

[0087] r kl =|f k -fl |

[0088] Among them, r kl represents the state distance between target k and target l, f k represents the state function of target k, f l Represents the state function of target l.

[0089] It should be noted that in step S14, first, the target state distance is calculated according to the target state space. The target state distance refers to the difference or distance between two target state functions in the state space. This distance can be defined in a variety of ways. For example, the present invention selects the Euclidean distance. Of course, depending on the application scenario and user needs, the Manhattan distance or other distance metrics suitable for a specific application scenario can also be selected. The present invention does not limit this. In this step, the absolute value distance formula r is used. kl =|f k -f l | to calculate the state distance between target k and target l, where f k and f l Represent the state functions of target k and target l respectively. The calculation of the target state distance requires quantifying the state function of each target, which involves extracting or calculating the state parameters of each target from the state space, including position, speed, etc. The calculation process includes comparing each dimension of the state vector, and then deriving the distance between the two states according to the selected distance formula. After obtaining the target state distance, a matrix is ​​constructed according to the target state distance to obtain the mutual relationship matrix. The mutual relationship matrix is ​​a square matrix whose size is equal to the square of the number of targets. Each element r in the matrix kl Represents the state distance between two targets. This matrix can be used to represent the relationship, proximity, interaction, etc. between targets. The relationship matrix is ​​of great use in multi-target tracking. It can be used for data association to help the algorithm determine the correspondence between observed data and target states. In addition, the relationship matrix can also be used to optimize the prediction and update steps of the target, and improve the tracking accuracy by considering the interaction between targets. In complex tracking scenarios, the relationship matrix provides a way to quantify the relationship between targets, which helps to deal with problems such as target occlusion and intersection. In this way, the relationship matrix enhances the robustness and adaptability of the multi-target tracking algorithm.

[0090] In step S15, a maximum expectation iteration operation is performed according to the target state space and the mutual relationship matrix, and when it is determined that the number of iterations is greater than a preset value, an estimated state is output, including:

[0091] According to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data;

[0092] According to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space, a posterior probability maximization operation is performed to obtain the optimal model state parameters;

[0093] Performing a state estimation operation according to the target state function, the observation data and the optimal model state parameters to obtain an estimated state;

[0094] According to the estimated state, construct a mutual relationship matrix to obtain a second relationship matrix;

[0095] Perform matrix distance calculation according to the second relationship matrix and the mutual relationship matrix to obtain updated distance data;

[0096] When the updated distance data is less than a set threshold or the number of iterations is greater than a preset iteration threshold, the iteration ends.

[0097] It should be noted that in step S15, first, according to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data. Dirichlet distribution is a multivariate probability distribution, which is commonly used in parameter modeling of multinomial distribution. It can provide a probability weight for the state function of each target, and these weight data are obtained by random sampling of Dirichlet distribution. Then, according to the actual parameter weight, observation data, estimated parameter weight data and target state space, the posterior probability maximization operation is performed to obtain the optimal model state parameters. This step is to adjust the model parameters by maximizing the posterior probability so that the model can best explain the observation data, which involves a complex optimization algorithm. Exemplarily, the present invention adopts the gradient descent method to implement the optimization process. Of course, according to the application scenario and user needs, Newton's method or other methods can also be selected, and the present invention does not limit this. Then, according to the target state function, observation data and optimal model state parameters, a state estimation operation is performed to obtain an estimated state. Exemplarily, the present invention adopts the Kalman filter method for state estimation. Of course, according to the application scenario and user needs, other state estimation operations can also be selected, and the present invention does not limit this. This step is to use the updated model parameters to predict the state of each target at the current time step, including the position, speed, etc. of the target. Afterwards, the mutual relationship matrix is ​​constructed according to the estimated state, and the construction method is the same as the above mutual relationship matrix construction method to obtain the second relationship matrix. This matrix is ​​similar to the original mutual relationship matrix, but based on the updated target state, it reflects the new relationship and interaction between the targets. Finally, according to the second relationship matrix and the mutual relationship matrix, the matrix distance calculation is performed to obtain the updated distance data. This distance data is used to measure the degree of change of the mutual relationship matrix between two iterations. When the change is small, it means that the model has stabilized and the iteration can be terminated. When the updated distance data is less than the set threshold or the number of iterations is greater than the preset iteration threshold, the iteration is terminated. Exemplarily, the present invention sets the exit iteration threshold to 100 and the number of iterations to 5. Of course, other thresholds or other numbers of iterations can be selected according to different application scenarios and user needs, and the present invention does not limit this. This condition ensures that the algorithm can stop after reaching a certain accuracy or after a sufficient number of iterations, thereby avoiding unnecessary waste of computing resources. In this way, the maximum expectation iteration operation can gradually improve the accuracy of multi-target tracking and ultimately output a stable and accurate target state estimate.

[0098] It is worth noting that, according to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data, including:

[0099] According to the target state space, a target frequency statistical operation is performed to obtain target occurrence frequency data;

[0100] Perform a weight calculation operation according to the target occurrence number data, the preset Dirichlet hyperparameters and the dimension of the target state function to obtain estimated parameter weight data;

[0101] The weight calculation formula is as follows:

[0102]

[0103] Among them, n k =[n k (1),n k (2),…n k (d)…,n k (D)] T represents the number of times target k appears in the state space, n k (d) represents the number of times the target k appears in the dth state space of the state space, represents the total number of occurrences of all targets, M is the total number of state spaces, D is the dimension of the target state function f, α is the hyperparameter of the Dirichlet distribution, and θ kt Represents the parameter weight of target k at time t.

[0104] It should be noted that the process of performing Dirichlet distribution calculation based on the target state space and obtaining estimated parameter weight data involves two key operations: target frequency statistics and weight calculation. First, the target frequency statistics operation involves counting the number of times the target appears in the state space, which is accomplished by analyzing the distribution of the target state function in each state space to obtain the target occurrence data. Then, the weight calculation operation is performed using the target occurrence data, the preset Dirichlet hyperparameters, and the dimension of the target state function. In the weight calculation formula, n k =[n k (1),n k (2),…,n k (D)] T represents the number of times target k appears in the state space, n k (d) represents the number of times the target k appears in the dth state space of the state space, represents the total number of occurrences of all targets, M is the total number of state spaces, D is the dimension of the target state function f, α is the hyperparameter of the Dirichlet distribution, and θ kt represents the parameter weight of target k at time t. This calculation process adjusts the weight of each target state so that targets that appear more frequently in the state space receive higher weights and thus receive more attention in the subsequent tracking process. This weight distribution helps improve the tracking algorithm's sensitivity and accuracy to target state changes, especially when target states change frequently or when interactions between targets are complex.

[0105] It is worth noting that, according to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space, the posterior probability maximization operation is performed to obtain the optimal model state parameters, including:

[0106] Constructing a state posterior probability formula according to the estimated parameter weight data and the target state space;

[0107] According to the state posterior probability formula, a maximization solution operation is performed to obtain the optimal model state parameters;

[0108] The state posterior probability formula is as follows:

[0109]

[0110] Among them, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the t-th iteration, θ represents the model parameters, θ (t) represents the parameter estimate of the tth iteration, N is the number of samples, z n is the nth target state, p(z n |x n ,θ (t) ) is the posterior probability of the nth target state based on the model parameters at the tth iteration, p(x n ,z n |θ) is the complete data likelihood function, x n represents observation data;

[0111] The maximization solution formula is as follows:

[0112]

[0113] in, represents the optimal model state parameter obtained at the tth iteration, represents the selection of θ that maximizes Q, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the tth iteration.

[0114] It should be noted that in the multi-target tracking method, the posterior probability maximization operation is a key step, which involves constructing a state posterior probability formula and performing a maximization solution operation to obtain the optimal model state parameters. This process first requires constructing a state posterior probability formula based on the estimated parameter weight data and the target state space. In the state posterior probability formula, Q(θ|θ (t) ) represents the state posterior probability, θ represents the model parameter, θ (t) represents the parameter estimate of the tth iteration, N is the number of samples, z n is the nth target state, p(zn |x n ,θ (t) ) is the posterior probability of the nth target state based on the model parameters at the tth iteration, p(x n ,z n |θ) is the complete data likelihood function, x n Represents the observed data. After constructing the state posterior probability formula, the maximization operation is performed to find the state that maximizes Q(θ|θ (t) )’s model parameters θ.

[0115] In the maximization solution formula, represents the optimal model state parameter obtained at the tth iteration, represents the selection of θ that maximizes Q, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters of the t-th iteration. This maximization operation involves a complex optimization algorithm. Exemplarily, the present invention selects the gradient descent method to implement the optimization process. Of course, depending on the application scenario and user needs, Newton's method or other numerical optimization methods can also be selected, and the present invention is not limited to this. The optimization algorithm iteratively adjusts the model parameters to improve the fit of the model to the observed data. In the context of multi-target tracking, this process helps to accurately estimate the state of each target, including their position, speed, etc., thereby improving the accuracy and robustness of tracking. In this way, the posterior probability maximization operation provides an effective mechanism for the multi-target tracking algorithm to optimize the estimation of the target state in the presence of uncertainty and noise.

[0116] It is worth noting that, according to the target state function, the observed data and the optimal model state parameters, a state estimation operation is performed to obtain an estimated state, including:

[0117] Calculating the posterior probability according to the target state function and the optimal model state parameters to obtain posterior probability data;

[0118] Performing an update operation according to the posterior probability data and the observation data to obtain an estimated state;

[0119] The update calculation formula is as follows:

[0120]

[0121] in, is the updated value of the state function of the kth target in the t+1th iteration, γ(z ik ) is the posterior probability that the i-th data point belongs to the k-th state, z ik represents the i-th data point belonging to the k-th state, x iis the observed data, and n is the total number of observed data.

[0122] It should be noted that in the multi-target tracking method, the state estimation operation is a key step, which involves the calculation and update operation of the posterior probability to obtain the estimated state. First, the posterior probability is calculated according to the target state function and the optimal model state parameters. This step uses Bayes' theorem, combined with the observed data and model parameters, to calculate the posterior probability of each data point belonging to each target state and obtain the posterior probability data.

[0123] The calculation formula for the posterior probability data is as follows:

[0124]

[0125] In the calculation formula of the posterior probability data, γ(z ik ) is the posterior probability that the i-th data point belongs to the k-th state, z ik represents the i-th data point belonging to the k-th state, x i is the observation data, is the optimal model state parameter obtained in the tth iteration. After obtaining the posterior probability data, an update operation is performed based on the posterior probability data and the observation data to obtain the estimated state. The purpose of the update operation is to combine the observation data with the posterior probability to estimate the state of each target at the current moment. In the update calculation formula, is the updated value of the state function of the kth target in the t+1th iteration, γ(z ik ) is the posterior probability that the i-th data point belongs to the k-th state, x i is the observation data, and n is the total number of observation data. This update formula is essentially a weighted average of the observation data of each target state, with the weight given by the posterior probability, to obtain the estimated state of each target. This state estimation method allows the algorithm to accurately estimate the target state while taking into account the interactions and uncertainties between targets. In the context of multi-target tracking, this state estimation operation is crucial to improve tracking accuracy and robustness, especially in complex situations such as occlusion, intersection or sensor noise between targets. In this way, the state estimation operation provides an effective mechanism for multi-target tracking algorithms to optimize the estimation of target states in the presence of uncertainty and noise.

[0126] It is worth noting that, according to the second relationship matrix and the mutual relationship matrix, matrix distance calculation is performed to obtain updated distance data, including:

[0127] Performing a subtraction operation on the second relationship matrix and the mutual relationship matrix to obtain a relationship difference matrix;

[0128] According to the relationship difference matrix, a negative number is replaced by a positive operation to obtain a relationship absolute difference matrix;

[0129] According to the relationship absolute difference matrix, elements are summed to obtain updated distance data.

[0130] It should be noted that in the multi-target tracking method, matrix distance calculation is a key step for evaluating the degree of change in the mutual relationship between targets in two consecutive iterations. This process first involves subtracting the second relationship matrix and the mutual relationship matrix to obtain a relationship difference matrix. The second relationship matrix is ​​constructed based on the updated target state, while the mutual relationship matrix is ​​constructed based on the target state of the previous iteration. The purpose of the subtraction operation is to quantify the impact of the change in the target state between two iterations on the mutual relationship. Next, according to the relationship difference matrix, a negative positive operation is performed to obtain a relationship absolute difference matrix. This step ensures that all differences are non-negative, so that the magnitude of the state change can be accurately reflected without considering the direction of the change. Then, according to the relationship absolute difference matrix, the elements are summed to obtain the updated distance data. This summation operation provides an overall metric for evaluating the sum of the changes in the mutual relationships between all targets. The updated distance data is used to determine whether the algorithm has converged or whether further iterations are required. When it is determined that the updated distance data is less than the set threshold, this indicates that the change in the target state is small enough, and the tracking algorithm can be considered to have converged, and the iteration can be terminated at this time. When the number of iterations is determined to be greater than the preset iteration threshold, the iteration will be terminated even if the updated distance data does not reach the threshold to avoid unnecessary calculations. This matrix distance calculation process provides a method for multi-target tracking algorithms to monitor and evaluate changes in target states. By comparing the relationship matrices between consecutive iterations, the algorithm can adjust its parameters more flexibly to adapt to dynamic changes in target states. This approach helps improve tracking accuracy and robustness, especially when the relationships between targets are complex or the target states change frequently. In this way, matrix distance calculation provides an effective mechanism for multi-target tracking algorithms to optimize the estimation and update of target states in the presence of uncertainty and noise.

[0131] In summary, the present invention provides a multi-target tracking method, characterized in that it includes: obtaining the number of tracked targets, target state parameters, observation data and actual parameter weights; constructing a target state function according to the number of tracked targets and the target state parameters to obtain a target state function; constructing a state space according to the target state function to obtain a target state space; constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix; performing a maximum expected iteration operation according to the target state space and the mutual relationship matrix, and outputting an estimated state when it is determined that the number of iterations is greater than a preset value.

[0132] The multi-target tracking method provided by the present invention realizes the accurate construction of the target state function by comprehensively considering the number of tracked targets, target state parameters, observation data and actual parameter weights. This method first constructs the target state function according to the number of tracked targets and the target state parameters, and defines a detailed state model for each target. Then, based on the target state function, the state space is further constructed to provide a comprehensive description of the state change of each target. In addition, the present invention quantifies the mutual influence between different targets by constructing a mutual relationship matrix, which is crucial in multi-target tracking. The introduction of the mutual relationship matrix enables the algorithm to more accurately simulate the dynamic relationship between targets, especially when the targets are close or blocked. Finally, through the maximum expectation iteration operation, that is, the EM algorithm, the present invention continuously optimizes the estimation of the target state. The iterative process of the EM algorithm allows the algorithm to update the model parameters according to the current best estimate in each iteration, thereby gradually improving the accuracy of the estimation. When the number of iterations reaches a preset threshold, the algorithm outputs the final estimated state, which reflects the most likely position and state of the target during the tracking process. Therefore, the present invention significantly improves the accuracy of multi-target tracking by accurately constructing the target state function, state space and mutual relationship matrix, combined with the iterative optimization of the EM algorithm. This approach can not only handle the mutual influence between targets, but also adapt to the dynamic changes of target states, thereby providing more accurate and reliable tracking results in complex tracking scenarios.

[0133] Reference Figure 2 , an embodiment of the present invention provides a multi-target tracking system, comprising:

[0134] Input module, used to obtain the number of tracked targets, target state parameters and actual parameter weights;

[0135] A state function construction module is used to construct a target state function according to the number of tracking targets and the target state parameters to obtain a target state function;

[0136] A state space construction module, used to construct the state space according to the target state function to obtain the target state space;

[0137] A mutual relationship construction module, used for constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix;

[0138] The output module is used to perform a maximum expectation iteration operation according to the target state space and the mutual relationship matrix, and output an estimated state when it is determined that the number of iterations is greater than a preset value.

[0139] It should be noted that a multi-target tracking device provided in an embodiment of the present invention is used to execute all process steps of a multi-target tracking method in the above embodiment, and the working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0140] The embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned multi-target tracking method embodiments are implemented, for example Figure 1 Alternatively, the processor implements the functions of each module / unit in the above-mentioned device embodiments when executing the computer program.

[0141] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the electronic device.

[0142] The electronic device may be a computing device such as a desktop computer, a notebook, a PDA, and a smart tablet. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. The electronic device may include more or fewer components than the above components, or may combine certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0143] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the electronic device, and uses various interfaces and lines to connect various parts of the entire electronic device.

[0144] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the electronic device by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0145] Wherein, if the module / unit integrated in the electronic device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0146] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0147] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-target tracking method, characterized in that: Executed by a computer, including: Obtain the number of tracked targets, target state parameters, observation data and actual parameter weights; According to the number of tracking targets and the target state parameters, a target state function is constructed to obtain a target state function; According to the target state function, a state space is constructed to obtain a target state space; According to the target state space, a mutual relationship matrix is ​​constructed to obtain a mutual relationship matrix; According to the target state space and the mutual relationship matrix, a maximum expectation iteration operation is performed, and when it is determined that the number of iterations is greater than a preset value, an estimated state is output.

2. A multi-target tracking method according to claim 1, characterized in that: According to the number of tracking targets and the target state parameters, a target state function is constructed to obtain a target state function, including: Performing a normalization operation according to the number of tracked targets and the target state parameter to obtain a normalized target state; According to the normalized target state and the preset time interval, time division is performed to obtain a target state function.

3. A multi-target tracking method according to claim 1, characterized in that: According to the target state function, the state space is constructed to obtain the target state space, including: According to the target state function, a time alignment operation is performed to obtain an aligned target state function; According to the alignment target state function, the states are integrated at the same time to obtain the target state space.

4. A multi-target tracking method according to claim 1, characterized in that: According to the target state space, a mutual relationship matrix is ​​constructed to obtain a mutual relationship matrix, including: Calculating the target state distance according to the target state space to obtain the target state distance; According to the target state distance, a matrix is ​​constructed to obtain a mutual relationship matrix; The formula for calculating the target state distance is as follows: r kl =f k -f l Among them, r kl represents the state distance between target k and target l, f k represents the state function of target k, f l Represents the state function of target l.

5. A multi-target tracking method according to claim 1, characterized in that: According to the target state space and the mutual relationship matrix, a maximum expectation iteration operation is performed, and when it is determined that the number of iterations is greater than a preset value, an estimated state is output, including: According to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data; According to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space, a posterior probability maximization operation is performed to obtain the optimal model state parameters; Performing a state estimation operation according to the target state function, the observation data and the optimal model state parameters to obtain an estimated state; According to the estimated state, construct a mutual relationship matrix to obtain a second relationship matrix; Perform matrix distance calculation according to the second relationship matrix and the mutual relationship matrix to obtain updated distance data; When the updated distance data is less than a set threshold or the number of iterations is greater than a preset iteration threshold, the iteration ends.

6. A multi-target tracking method according to claim 5, characterized in that: According to the target state space, Dirichlet distribution calculation is performed to obtain estimated parameter weight data, including: According to the target state space, a target frequency statistical operation is performed to obtain target occurrence frequency data; Perform a weight calculation operation according to the target occurrence number data, the preset Dirichlet hyperparameters and the dimension of the target state function to obtain estimated parameter weight data; The weight calculation formula is as follows: Among them, n k =[n k (1),n k (2),…n k (d)…,n k (D)] T represents the number of times target k appears in the state space, n k (d) represents the number of times the target k appears in the dth state space of the state space, represents the total number of occurrences of all targets, M is the total number of state spaces, D is the dimension of the target state function f, α is the hyperparameter of the Dirichlet distribution, and θ kt Represents the parameter weight of target k at time t.

7. A multi-target tracking method according to claim 5, characterized in that: According to the actual parameter weight, the observed data, the estimated parameter weight data and the target state space, a posterior probability maximization operation is performed to obtain the optimal model state parameters, including: Constructing a state posterior probability formula according to the estimated parameter weight data and the target state space; According to the state posterior probability formula, a maximization solution operation is performed to obtain the optimal model state parameters; The state posterior probability formula is as follows: Among them, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the t-th iteration, θ represents the model parameters, θ (t) represents the parameter estimate of the tth iteration, N is the number of samples, z n is the nth target state, p(z n |x n ,θ (t) is the posterior probability of the nth target state based on the model parameters at the tth iteration, p(x n ,z n |θ) is the complete data likelihood function, x n represents observation data; The maximization solution formula is as follows: in, represents the optimal model state parameter obtained at the tth iteration, represents the selection of θ that maximizes Q, Q(θ|θ (t) ) represents the state posterior probability based on the model parameters at the tth iteration.

8. A multi-target tracking method according to claim 5, characterized in that: Performing a state estimation operation according to the target state function, the observation data and the optimal model state parameter to obtain an estimated state includes: Calculating the posterior probability according to the target state function and the optimal model state parameters to obtain posterior probability data; Performing an update operation according to the posterior probability data and the observation data to obtain an estimated state; The update calculation formula is as follows: in, is the updated value of the state function of the kth target in the t+1th iteration, γ(z ik ) is the posterior probability that the i-th data point belongs to the k-th state, z ik represents the i-th data point belonging to the k-th state, x i is the observed data, and n is the total number of observed data.

9. A multi-target tracking method according to claim 5, characterized in that: Performing matrix distance calculation according to the second relationship matrix and the mutual relationship matrix to obtain updated distance data includes: Performing a subtraction operation on the second relationship matrix and the mutual relationship matrix to obtain a relationship difference matrix; According to the relationship difference matrix, a negative number is replaced by a positive operation to obtain a relationship absolute difference matrix; According to the relationship absolute difference matrix, elements are summed to obtain updated distance data.

10. A multi-target tracking system, characterized in that: include: Input module, used to obtain the number of tracked targets, target state parameters and actual parameter weights; A state function construction module is used to construct a target state function according to the number of tracking targets and the target state parameters to obtain a target state function; A state space construction module, used to construct the state space according to the target state function to obtain the target state space; A mutual relationship construction module, used for constructing a mutual relationship matrix according to the target state space to obtain a mutual relationship matrix; The output module is used to perform a maximum expectation iteration operation according to the target state space and the mutual relationship matrix, and output an estimated state when it is determined that the number of iterations is greater than a preset value.

Images