Multi-target tracking method and system

Abstract

The invention discloses a multi-target tracking method and system. The method comprises the following steps: a state prediction step: predicting a target state and covariance at a current moment based on target state estimation at a previous moment; a factor graph construction step: a bipartite factor graph model containing two types of nodes representing a target and measurement and an edge between the nodes is established, and the edge represents an association possibility weight; a message iteration step of performing multiple rounds of message passing on the bipartite factor graph, in at least one round, applying a damping factor to an outgoing message for weighted smoothing, and / or applying a scaling factor to an incoming message for confidence correction; and a probability calculation and state updating step: calculating an edge association probability between the target and the measurement according to an iteration result, and updating a target state and a covariance according to the edge association probability. According to the method, the problems that a traditional method under the dense clutter and target crossing scene is too high in calculation complexity, track condensation is prone to occurring, and estimation is excessively confident due to a factor graph loop are solved, and tracking precision and stability are improved.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and target tracking technology, and more specifically to a method for tracking target states in dense clutter and multi-target intersection scenarios by utilizing cyclic belief propagation and joint probability data association. Background Technology

[0002] Multi-target tracking technology is widely used in radar surveillance, autonomous driving, and air traffic control. Its core task is to jointly estimate the number of targets and their motion states (such as position and velocity) over time from sensor measurement data containing noise, clutter, and false alarms. The most critical and challenging aspect of multi-target tracking is data correlation, i.e., determining which measurement originates from which target, or whether it originates from clutter background, when the measurement source is uncertain. Due to sensor resolution limitations and environmental interference, the correspondence between measurements and targets is often highly ambiguous; incorrect correlation will directly lead to filter divergence and track loss.

[0003] For data association problems, the Global Nearest Neighbor (GNN) algorithm is a typical "hard-decision" approach that seeks a globally optimal single association hypothesis by minimizing the total distance or maximizing the likelihood function. While GNN is relatively simple to compute, it is extremely vulnerable to dense clutter or close proximity of targets, easily leading to false associations. In contrast, the Joint Probabilistic Data Association Algorithm employs a "soft association" strategy based on Bayesian inference. Instead of making either-or decisions, it calculates the marginal association probability of each measurement originating from each target and updates the target state using a weighted sum of all feasible association hypotheses.

[0004] To reduce computational complexity while maintaining accuracy, the factor graph-based belief propagation algorithm (also known as the sum-product algorithm) has been introduced into the field of data association. The belief propagation algorithm models the data association problem as a graph model containing variable nodes (representing association states) and factor nodes (representing constraints), approximating the probability of marginal associations by iteratively passing probability "messages" between nodes. Compared to the combinatorial enumeration of traditional methods, the belief propagation algorithm leverages the conditional independence of the statistical model to significantly reduce computational complexity and greatly improve the algorithm's scalability.

[0005] While joint probabilistic data association algorithms form the foundation for multi-target tracking, existing technologies suffer from significant and insurmountable drawbacks when dealing with large-scale, dense target scenarios. First, the computational complexity of traditional algorithms increases exponentially. As relevant research has pointed out, accurately calculating the posterior probability of the core of this algorithm—the joint association event—is mathematically equivalent to calculating the product of matrices, a notoriously difficult counting problem in computational complexity theory. With the increase in the number of targets and sensor measurements within the field of view, the number of feasible joint association hypotheses expands dramatically, causing the computational load to rapidly exceed the processing capacity of real-time systems, making it unsuitable for applications in high-density, complex environments.

[0006] Secondly, when dealing with spatially probabilistic targets, joint probabilistic data association algorithms inherently exhibit a "track aggregation" phenomenon. When two or more targets are very close in space (e.g., moving in parallel or crossing), the estimation criterion based on minimum mean square error will perform a weighted average of the shared measurements of neighboring targets, causing the state estimates to tend towards the geometric center of the measurement group rather than their respective true positions. This results in the originally separate tracks attracting each other and eventually merging, leading to the loss of target identity.

[0007] Finally, while factor graph-based belief propagation algorithms can reduce computational complexity, standard algorithms have limitations when dealing with graph structures specific to data association. Since factor graphs of data associations often contain closed-loop structures, directly applying standard belief propagation algorithms can lead to cyclical feedback of information within these loops, producing an "echo" effect. This positive feedback effect causes the algorithm to calculate marginal probability distributions with a non-physical "overconfidence," artificially reducing the estimated variance and severely underestimating the uncertainty of the association. This can easily lead to incorrect association decisions in scenarios with low signal-to-noise ratios or dense clutter. Summary of the Invention

[0008] The present invention aims to address the aforementioned deficiencies in the prior art. Specifically, the technical problem to be solved by the present invention is: how to overcome the problems of high computational complexity, easy track aggregation, and "overconfidence" in estimation caused by factor graph loops in multi-target tracking methods based on joint probability data association in dense clutter and multi-target intersection scenarios.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A multi-target tracking method includes: a state prediction step, comprising predicting the target state and predicted covariance at the current time based on the target state estimate of the previous time step; a factor graph construction step, comprising establishing a bipartite factor graph model, wherein the bipartite factor graph model includes a first type of node representing the target, a second type of node representing the measurement, and an edge connecting the first type of node and the second type of node, wherein the edge is used to represent the association probability weight between the target and the measurement; a message iteration step, comprising performing multiple rounds of message passing between the first type of node and the second type of node on the bipartite factor graph model, wherein in at least one round of message passing, a damping factor is applied to the outgoing message for weighted smoothing, and / or a scaling factor is applied to the incoming message for confidence correction; and a probability calculation and state update step, comprising calculating the edge association probability between the target and the measurement based on the converged result of the message iteration step, and updating the target state estimate and estimated covariance at the current time based on the edge association probability.

[0010] In some embodiments, at least one of the following technical means is also included: Before the factor graph construction step, the method further includes a gate screening step, which includes calculating the statistical distance between the measurement and the target predicted value, and screening out effective association pairs according to a preset threshold. The edges in the bipartite factor graph model are established only based on the effective association pairs.

[0011] The message iteration steps include: a first message calculation step, which includes calculating the message sent from a first type of node to a second type of node based on the initial association weight of the target and measurement and other messages passed in from the second type of node, and introducing the scaling factor in the calculation; a first message update step, which includes weighting and fusing the message obtained from the first message calculation step with the corresponding historical message from the previous iteration using the damping factor to generate the final message sent to the second type of node; a second message calculation step, which includes calculating the message sent from a second type of node to a first type of node based on the initial association weight of the target and measurement and other messages passed in from the first type of node, and introducing the scaling factor in the calculation; and a second message update step, which includes weighting and fusing the message obtained from the second message calculation step with the corresponding historical message from the previous iteration using the damping factor to generate the final message sent to the first type of node.

[0012] The calculations in the first message calculation step follow the following mathematical relationships: ,in For the target in the p-th iteration Send to Measurement Intermediate message, As the initial association weight, For other measurements in the previous round Send to target The news The scaling factor is mentioned.

[0013] The calculations in the second message calculation step follow the following mathematical relationships: ,in Measurement in the p-th iteration Send to target Intermediate message, As the initial association weight, For other objectives in the current round Send to Measurement The news The scaling factor reflects the constraint that "a measurement can originate from at most one target".

[0014] In the probability calculation and state update steps, the scaling factor is introduced to correct the probability distribution when calculating the edge association probability. Its calculation follows the following mathematical relationship: ,in For the goal With measurement Edge association probability, As the initial association weight, For all measurements after iterative convergence Send to target The news, The scaling factor is mentioned.

[0015] The state prediction step and the probability calculation and state update step are implemented based on the Kalman filter framework.

[0016] The damping factor has a value range of [0, 1), and the scaling factor has a value range of (0, 1).

[0017] The termination condition for the message iteration step is: the maximum difference between the message matrices generated by two consecutive iterations is less than a preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations.

[0018] The present invention also adopts the following technical solutions: A signal processing system for multi-target tracking includes: a memory for storing a computer program; and a processor for executing the computer program in the memory to implement the multi-target tracking method as described in any one of the preceding embodiments.

[0019] The present invention also adopts the following technical solutions: A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the multi-target tracking method as described in any of the preceding claims.

[0020] Furthermore, A motor vehicle includes the aforementioned signal processing system for realizing environmental perception and multi-target tracking in autonomous driving.

[0021] Furthermore, An air traffic control system includes the aforementioned signal processing system for monitoring and tracking the status of aircraft in the airspace.

[0022] Furthermore, An unmanned aerial vehicle (UAV) includes the aforementioned signal processing system for achieving multi-target tracking in autonomous navigation and obstacle avoidance.

[0023] The present invention also adopts the following technical solutions: A method for storing a bit stream of tracking data generated by a specific multi-target tracking method, comprising performing the multi-target tracking method to generate a bit stream containing target state information; and storing the bit stream.

[0024] The present invention also adopts the following technical solutions: A method for transmitting a tracking data bitstream generated by a specific multi-target tracking method, comprising performing the multi-target tracking method to generate a bitstream containing target state information; and transmitting the bitstream.

[0025] The present invention also adopts the following technical solutions: A computer-readable storage medium storing a computer program / instructions and a bit stream thereon, wherein the computer program / instructions, when executed by a processor, implement the multi-target tracking method described above to generate the bit stream, the bit stream containing state estimation information of multiple targets.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively suppresses positive feedback and "echo" effects caused by loops in the factor graph by introducing a damping factor to weighted smooth outgoing messages and / or a scaling factor to correct the confidence of incoming messages during the message passing iteration process of a bipartite factor graph model. The introduction of the damping factor increases the numerical inertia of the iteration process, preventing drastic oscillations in message updates and ensuring the convergence stability of the algorithm in complex association scenarios. The introduction of the scaling factor adjusts the entropy of the message or probability distribution, correcting the "overconfidence" bias caused by loops, making the calculated edge association probabilities more accurately reflect the uncertainty of the association. These two measures work synergistically to make the final edge association probabilities more accurate and reliable. Target state updates based on these more accurate edge association probabilities fundamentally alleviate the "track aggregation" problem caused by probability calculation errors in traditional joint probabilistic data association algorithms, significantly improving tracking accuracy and track maintenance capabilities in environments with multiple target intersections, parallel movements, and dense clutter. Meanwhile, based on the factor graph model and message passing mechanism, this invention avoids the combined enumeration of joint events in traditional methods, reducing the computational complexity from exponential to polynomial (usually linear or quadratic). This greatly improves the efficiency of the algorithm while ensuring high accuracy, enabling it to meet the needs of high real-time application scenarios.

[0027] Furthermore, by pre-eliminating invalid association pairs through the gate filtering step, the number of edges in the bipartite factor graph is reduced, thereby further reducing the computational burden of message iteration.

[0028] Furthermore, by clarifying the mathematical relationships and steps of message iteration, the feasibility and reproducibility of the algorithm are ensured.

[0029] Furthermore, by introducing a scaling factor in the marginal probability calculation, the consistency between the probability output and the iterative process correction logic is ensured.

[0030] Furthermore, by implementing state prediction and updating based on a mature Kalman filter framework, the stability and reliability of the tracking process are ensured.

[0031] Furthermore, by limiting the range of values ​​for the damping factor and scaling factor, clear guidance is provided for parameter adjustment, which is beneficial for algorithm performance optimization.

[0032] Furthermore, by setting explicit iteration termination conditions, infinite loops are avoided, ensuring the practicality of the algorithm.

[0033] Furthermore, by providing corresponding signal processing systems, storage media, and various application devices (such as motor vehicles, drones, etc.), the beneficial effects at the method level are extended to the system and application levels, thus broadening the scope of protection and application value of the present invention.

[0034] Furthermore, by defining the method for generating and storing / transmitting the trace data bitstream, the intermediate data products generated by the method of the present invention and their processing flow are protected. Attached Figure Description

[0035] Figure 1 This is a flowchart of a multi-target tracking method based on cyclic belief propagation and joint probability data association provided in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the structure of the target-measurement dichotomy factorial graph constructed in one embodiment of the present invention.

[0037] Figure 3 This is a detailed flowchart of the iterative steps of performing cyclic belief propagation messages in one embodiment of the present invention.

[0038] Figure 4 This is a comparison curve of the expected optimal submode assignment distance (E-OSPA) of the proposed method (CBP-JPDA) and the traditional method (LBP-JPDA) as a function of time steps in a simulation experiment. Detailed Implementation

[0039] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0040] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to that other component. Furthermore, a connection can be used for both fixing and circuit / signal connectivity.

[0041] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0042] Furthermore, 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0043] The basic concept of this invention is as follows: This invention proposes a multi-target tracking method based on cyclic belief propagation and joint probability data association. The method first predicts the motion state of multiple targets using a Kalman filter framework; then, it constructs a bipartite factor graph model with targets and measurements as nodes, and uses a parameterized cyclic belief propagation algorithm for iterative message passing. This algorithm actively cancels redundant information echoes caused by loops by weighting and correcting the feedback messages between factor nodes and variable nodes during the iteration process, thus accurately approximating edge association probabilities with low computational complexity. Finally, it uses the converged association probabilities to update the target state and covariance of the joint probability data association, effectively solving the computational bottleneck and track aggregation problem under dense clutter.

[0044] The specific implementation steps of this invention are as follows: Step S1: Initialize parameters, build the model and predict the state In this step, the control parameters of the algorithm are set, a mathematical model of the target motion and observation is established, and the target state at the current moment is deduced using the posterior estimate of the previous moment.

[0045] Step S11: Set the core control parameters of the cyclic belief propagation algorithm to correct the belief distribution in subsequent message passing iterations. Specifically, define the damping factor. and scaling factor .

[0046] Damping factor Used to control the "inertia" during message update, its value is strictly limited to the interval [0,1) (typically, the value range is...). In scenarios with dense clutter or target intersections, where factor graph loops are easily formed, standard message passing is prone to numerical oscillations or even non-convergence. Introducing a damping factor... Subsequently, the new message calculation will adopt a weighted average of the "current calculated value" and the "previous round of historical messages" (i.e., This mechanism is based on fixed-point iteration theory. By introducing numerical inertia, it smooths the update amplitude of messages, effectively preventing algorithm divergence caused by positive feedback loops, and ensuring that the algorithm can stably converge to an effective marginal probability distribution.

[0047] scaling factor Also known as the reweighting parameter, it is used to combat the "overconfidence" effect in confidence levels caused by loops in the factor graph, and its value is set to (0,1). In factor graphs with loops, the same information is often calculated repeatedly, resulting in an overly sharp marginal probability distribution (too low entropy). Scaling factor Based on reweighted Bethe Free Energy, exponential scaling is performed on the likelihood function or aggregated message. This increases the distribution entropy.

[0048] The damping factor The scaling factor controls the smoothness of message updates. A value closer to 1 indicates a higher weight for historical messages, leading to more stable but potentially slower algorithm convergence; a value closer to 0 results in faster response to new messages but may cause oscillations. The sharpness used to adjust the confidence level; the smaller the value is (less than 1), the stronger the correction against 'overconfidence,' and the smoother the output probability distribution; a value of 1 is equivalent to the standard algorithm. The above parameters ( , The optimal value of is related to the characteristics of the specific application scenario, including but not limited to target density, clutter rate, and sensor accuracy. Those skilled in the art can optimize the parameters for specific scenarios through conventional simulation experiments (e.g., mesh search). Typical values ​​given in the embodiments of this specification ( =0.5, =0.8) is an effective value obtained after parameter tuning in a simulation environment [e.g., target number 8, clutter rate 70, detection probability 0.85], and can be used for initial settings in similar scenarios.

[0049] Step S12: Establish a state-space model that describes the dynamic changes in the target state and the measurement characteristics of the radar system, providing a benchmark for subsequent state prediction and updates.

[0050] First, a target transition equation is constructed to describe the state changes of the target in continuous or discrete time, in the form of:

[0051] in, Indicates the first The target state at each pulse moment (usually including components such as position and velocity). It is the state transition matrix. For process noise that follows a known Gaussian distribution, its covariance matrix is: .

[0052] Secondly, a radar measurement equation is constructed to describe the relationship between the target state and radar measurements, in the form of:

[0053] in, Indicates the first Measurement vector at each pulse moment It is a nonlinear mapping of measurement functions (such as distance, azimuth, and elevation). The measurement noise is a multivariate Gaussian distribution, and its covariance matrix is: .

[0054] Step S13: Use the Kalman filter prediction equation, combined with the transition matrix from step S12. Process noise covariance Based on the posterior state of the previous time step Calculate the predicted prior state value at the current time. and covariance :

[0055]

[0056] in, and These are the posterior state estimate and posterior covariance of the previous time step, respectively. This prediction result will serve as prior information input for subsequent data association and measurement updates.

[0057] Step S2: Construct the measurement correlation model and factor graph

[0058] The effective measurements are screened using the gate technique, the statistical distance between the measurements and the target is calculated, the initial likelihood weights describing the association probability weights are constructed, and a bipartite factor graph model is established for subsequent inference.

[0059] Step S21: Calculate the radar measurement at the current time With each target prediction measurement The Mahalanobis distance between them is used to eliminate invalid correlations that fall outside the gate based on a preset gate threshold, thereby reducing subsequent computational load. For the target and measurement An association is considered valid if the following conditions are met:

[0060] in, For the predictive measurement of the target, The new information covariance matrix, To correspond to the threshold probability The threshold, specifically, The value is the degree of freedom of the measurement vector dimension. The chi-square distribution at confidence level The quantiles below. In practical applications, they are usually set to... The value is between 0.95 and 0.99 to cover the vast majority of real-world measurements. For example, for common two-dimensional radar measurements ( If taken Then, by looking up the table, we can obtain... A verification matrix is ​​constructed based on the screening results. A matrix element of 1 indicates that the measurement falls within the confirmation gate of the target, while an element of 0 indicates that it is irrelevant.

[0061] Step S22: Calculate the initial association likelihood weights Based on the radar measurement equations and clutter model, the target is calculated. With effective measurement The initial correlation likelihood weights between them. For "measurement" Originating from the goal The likelihood weight of the association hypothesis is defined as follows:

[0062] For detection probability, clutter density, This is a Gaussian probability density function. The physical meaning of clutter density is the expected number of false measurements (clutter) per unit observation volume (or area). Determination methods typically include empirical methods and online estimation methods. The hypothesis that the target was not detected (missed detection) is denoted as... The weights are set as follows:

[0063] Step S23: Construct a target-measurement bipartite factorial graph model. In this step, based on the validation matrix generated in step S21 and the initial association likelihood weights calculated in step S22, a bipartite factorial graph is constructed to describe the multi-objective data association problem. Its topology is as follows: Figure 2 As shown. The left column in the figure represents variable nodes, corresponding to the various targets to be tracked (labeled as...). These are responsible for maintaining the probabilistic information of the target's association with different measurements; the column on the right side of the diagram represents factor nodes, corresponding to the radar measurements received at the current moment (labeled as...). The diagram illustrates the physical mutual exclusion constraint that "a measurement originates from at most one target." The connecting lines in the diagram represent valid correlation hypotheses after gate filtering, meaning that a hypothesis is only valid at the corresponding variable node if the corresponding element of the verification matrix in step S21 is 1. With factor nodes Connection edges are established between them. This gate-based sparse connection structure eliminates low-probability invalid associations, decoupling the globally complex association problem into a local message passing process. In addition, this step also requires initializing the bidirectional messages on all connection edges in the graph (usually set to unit values ​​or uniform distribution) to provide a computational starting point for the cyclic belief propagation iteration in the subsequent step S3.

[0064] Step S3: Perform iterative circular belief propagation messages.

[0065] Iterative message passing is performed on the constructed target-measurement bipartite factor graph to approximate the edge association probability. To address the non-convergence and "overconfidence" issues caused by factor graph loops, this invention adopts a step-by-step processing strategy: Steps S31 and S33 are message generation processes based on the standard belief propagation algorithm using Bayesian inference, responsible for calculating the intermediate information of the target node and measurement node under the current topology; Steps S32 and S34 are stability correction processes that introduce a damping mechanism, responsible for smoothing message oscillations using historical information to ensure the convergence of the algorithm in the looped graph; Step S35 checks the iterative convergence and obtains the final convergence value for subsequent edge probability calculation. Specifically, it includes the following five sub-steps: Step S31: Calculate the target node Send to measurement node intermediate message variables This step executes the "variable node update" logic in the standard belief propagation algorithm. Target node As a variable node, the aggregation comes from sources other than the current measurement. Feedback information from all other measurement nodes, except for the initial likelihood weights, is used to generate new correlation hypotheses. To combat redundant calculations caused by loops, a scaling factor is introduced on top of this standard procedure. The distribution is corrected. The calculation formula is:

[0066] in, For the goal With measurement The initial association likelihood weights; In the previous iteration, excluding measurement Other measurement nodes Feedback to the target The message. The "1" in the denominator represents the normalized weight term for targets that were not detected (missed detection).

[0067] Step S32: Apply damping factor Update the message sent from the target node to the measurement node. This step involves performing numerical stability correction on the standard message generated by S31. This is done using a preset damping factor. (range of values) The intermediate message variables are weighted and merged with the historical messages from the previous round. The update formula is:

[0068] Step S33: Calculate measurement nodes Send to target node intermediate message variables This step executes the factor node update logic in the standard belief propagation algorithm. Measurement node As a factor node, it is responsible for enforcing mutual exclusion constraints in data association (i.e., a measurement originates from at most one target). The scaling factor is also applied when calculating intermediate messages. Adjust the distribution:

[0069] in, In the previous iteration, the target was removed. Other target nodes Send to measurement The message. The summation term in the denominator reflects the total probability that the measure is occupied by other competing objectives.

[0070] Step S34: Apply damping factor Update the message sent from the measurement node to the target node. This step involves numerical stability correction of the standard feedback message generated by S33. The damping update is also performed using a weighted fusion method.

[0071] Step S35: Check the iteration convergence. Calculate the maximum absolute difference between the message matrix generated in the current iteration and the message matrix from the previous iteration, denoted as . .like Less than the preset convergence threshold (This value is usually set to a positive number much less than 1, such as...) ), or the current iteration number Reaching the preset maximum number of iterations (As a mandatory termination condition to prevent infinite loops, its value should be adjusted according to the target density and clutter complexity in the scene, usually set between 20 and 100), then it is determined that the algorithm has converged or reached the computational boundary, the iteration stops and the process proceeds to step S4 to calculate the edge probability; otherwise, let Return to step S31.

[0072] Step S4: Edge Association Probability Calculation and State Update

[0073] In this step, the edge association probability is calculated using the message obtained after the iterative convergence of cyclic belief propagation, and the weighted update of the target state is completed accordingly.

[0074] Step S41: Calculate the edge association probability and apply a scaling factor correction. Based on the final message after convergence in step S3, and combined with the initial likelihood weights calculated in step S2, calculate the target... With measurement Approximate marginal correlation probability To correct the "overconfidence" problem in confidence levels caused by factor graph loops, a scaling factor is applied here. The final confidence distribution is then adjusted for entropy. The calculation formula is as follows:

[0075] in The goal With measurement The initial association likelihood weights, At the end of the iteration, with the target All connected measurement nodes (including measurement) (Measurements of itself and other competitors) sent to the target The final feedback message It is a normalization constant to ensure the target All possible related events (including those related to each measurement) and the probability of missed detection The sum of 1 is 1.

[0076] Step S42: Utilize the calculated edge association probability The target state is updated according to the standard procedure of the joint probability data association algorithm. First, the target state is calculated. The combined information, i.e., the weighted sum of all valid measurement information:

[0077] Next, using Kalman gain Correct the state estimate and update the error covariance matrix. The covariance update must include the innovation scatter term caused by the uncertainty in data association. :

[0078]

[0079] in, This step not only updates the target location through weighted averaging, but also... This accurately reflects the increased uncertainty caused by the ambiguity of the association.

[0080] Example 1: In this embodiment, a multi-target tracking scenario is considered, in which eight moving targets exist simultaneously. In the simulation environment, the clutter rate is set to 70 (i.e., the average number of clutter measurements per time step), and the target detection probability is set to 0.85. The simulation duration is 100 time steps, and 50 Monte Carlo trials are conducted to ensure the reliability of the statistical results.

[0081] In this embodiment, the proposed Cyclic Belief Propagation Joint Probability Data Association (LBP-JPDA) method and the traditional LBP-JPDA method are used respectively for real-time estimation and tracking of multi-target states. Both methods use the exact same true trajectories, measurement data (including clutter), and motion and measurement models in the simulation; the only difference lies in the control parameter settings during message passing. The key parameter settings for the proposed method are as follows: loop correction parameters... Belief gain parameter The convergence tolerance of the proposed method and the LBP-JPDA method is... The maximum number of iterations is 100.

[0082] In this embodiment, the Expected Optimal Submode Allocation Distance (E-OSPA) is used as the evaluation metric, and the specific parameter is set to the order. (Corresponding to Euclidean distance metric) and cutoff distance (Used to limit the impact of base error on overall indicators).

[0083] Simulation results are as follows Figure 4 The figure shows a comparison of the proposed method with the traditional LBP-JPDA method in terms of multi-target tracking accuracy, using the Expected Optimal Submode Allocation (E-OSPA) distance as the core evaluation metric. The horizontal axis represents the time steps of the simulation experiment (from step 0 to step 100), and the vertical axis represents the E-OSPA distance. This metric can simultaneously quantify the target position estimation bias and the cardinality error (i.e., the number of missed or over-detected targets). The lower the value, the closer the tracking result is to the real scene, and the higher the accuracy. The blue solid line represents the traditional LBP-JPDA method as the benchmark, and the red solid line represents the improved method based on parameter correction proposed in this invention. The shaded areas around each line reflect the standard deviation of multiple Monte Carlo simulation experiments, representing the stability of the algorithm.

[0084] The evolution trend of the experimental curves clearly shows that although the two methods perform similarly in the initial stage of the track (0 to 10 steps), the method proposed in this invention (red line) exhibits a significant accuracy advantage over time. After entering the stable tracking stage, the red curve remains below the blue curve, stabilizing the E-OSPA error at around 4.2, while the error of the traditional method remains around 4.8. This result intuitively verifies the effectiveness of the technical improvement: by introducing damping and scaling factors in cyclic confidence propagation, the algorithm successfully combats the "overconfidence" effect of confidence caused by factor graph loops, actively offsetting the positive feedback echo of redundant information. This mechanism makes the calculation of edge correlation probability more accurate, thus effectively avoiding the track convergence or divergence problems common in traditional methods in multi-target intersection or dense clutter scenarios, significantly improving the estimation accuracy of target state.

[0085] Furthermore, comparing the widths of the shaded areas on the two curves reveals that this invention significantly reduces tracking errors without sacrificing algorithm stability. The red and blue shaded areas are roughly the same width, indicating that the method exhibits good robustness against random clutter and measurement noise. In summary, this figure strongly supports the core beneficial effect of this invention: by weighted correction of the message passing process while ensuring computational efficiency, it achieves superior and more robust multi-target tracking performance compared to traditional joint probabilistic data association methods.

[0086] It should be noted that the above embodiments and the specific parameter settings involved are only used to illustrate the effectiveness of the technical solution of the present invention and do not constitute a limitation on the scope of protection of the present invention. The scope of protection of the present invention should be determined by the contents of the appended claims. All equivalent structural substitutions, technical feature transformations, etc., made in accordance with this specification and drawings should be included within the scope of protection of the present invention.

[0087] In summary, this invention solves the computational complexity bottleneck in existing joint probabilistic data association methods, and demonstrates significant advantages in estimation accuracy and adaptability to complex environments, thus possessing important technical value and application prospects.

[0088] Experimental data:

[0089] In summary, the experimental data show that the average tracking error (E-OSPA) of the CBP-JPDA method proposed in this invention is [4.34 ± 0.11] throughout the entire 100 time steps of the simulation, which is significantly lower than that of the traditional LBP-JPDA method [4.72 ± 0.08]. Especially after the system enters the stable tracking phase (approximately 50 time steps later), the performance advantage of this method becomes more obvious and stable, with the error standard deviation consistently maintained at a low level, fully verifying the beneficial effects of this invention in improving tracking accuracy and robustness.

[0090] Example 2

[0091] This embodiment details the specific implementation of a multi-target tracking method.

[0092] The multi-target tracking method includes: a state prediction step, which includes predicting the target state prediction value and prediction covariance at the current time based on the target state estimate at the previous time step; a factor graph construction step, which includes establishing a bipartite factor graph model, wherein the bipartite factor graph model includes a first type of node representing the target, a second type of node representing the measurement, and an edge connecting the first type of node and the second type of node, wherein the edge is used to represent the probability weight of the association between the target and the measurement; a message iteration step, which includes performing multiple rounds of message passing between the first type of node and the second type of node on the bipartite factor graph model, wherein in at least one round of message passing, a damping factor is applied to the outgoing message for weighted smoothing, and / or a scaling factor is applied to the incoming message for confidence correction; and a probability calculation and state update step, which includes calculating the edge association probability between the target and the measurement based on the converged result of the message iteration step, and updating the target state estimate and estimated covariance at the current time step based on the edge association probability.

[0093] Explanation and correspondence: State prediction step: In this embodiment, this step is implemented using the prediction equation of the Kalman filter. Specifically, the target state estimate at the previous time step is the posterior state estimate of the Kalman filter, and the prediction at the current time step is calculated using the state transition matrix.

[0094] Bipartite factorial graph model: such as Figure 2As shown, this model is the core of constructing the relationships. The first type of node (variable node) is specified as a node representing each target, such as T1, T2, ..., Tm. The second type of node (factor node) is specified as a node representing each measurement, such as M1, M2, ..., Mn. An edge is specified as a line segment connecting the target and measurement nodes; its existence indicates that the measurement falls within the gate of the target, suggesting a possible association.

[0095] Message passing: This is specifically defined as messages that are iteratively computed between nodes in a factor graph, such as from a target... Send to Measurement News and from measurement Send to target News These messages are a representation of probabilistic information.

[0096] Damping factor: This is expressed as a numerical parameter, denoted as... Its value ranges from [0, 1), and it is used to perform a weighted average of the calculated new message and the previous round of historical messages, thus having a smoothing effect. For example =0.5.

[0097] Scaling factor: This is expressed as a numerical parameter, denoted as... Its value range is (0, 1], and it is used to exponentially scale the calculated message or probability, thereby correcting the confidence distribution. For example =0.8.

[0098] Marginal association probability: concretized as a probability value , indicating measurement Originating from the target The posterior probability.

[0099] Specific shortcomings addressed: This embodiment aims to solve the problem of "overconfidence" caused by information loop feedback when the standard belief propagation algorithm performs data association in a graph containing loop factors, as well as the problems of high computational complexity and easy track aggregation in traditional joint probabilistic data association algorithms.

[0100] Detailed technical solution: The system in this embodiment can be deployed on a computer or embedded system equipped with a processor (e.g., an Intel Xeon series CPU or an NVIDIA Tesla series GPU) and memory (e.g., DDR4 RAM). Measurement data acquired by sensors (such as radar) is input into the system. The method executes the following steps sequentially within each tracking cycle, and its overall flow is as follows: Figure 1 As shown: 1. State prediction: Based on the posterior state estimate of the previous time step k-1. _ (k-1|k-1) and error covariance P_ (k-1|k-1), using the Kalman filter prediction equation to calculate the prior state prediction at time k. _ (k|k-1) and prediction covariance P_ (k|k-1). The state transition matrix F and the process noise covariance Q need to be preset according to the target motion model.

[0101] 2. Gate filtering: For each target Calculate its predicted location (by _ (k|k-1) obtained with each measurement z_ The Mahalanobis distance between (k) is d_{ }^2. Set a gate threshold γ (for example, for two-dimensional measurements, take γ=9.21, corresponding to a 97.5% confidence level). If d_{ If}^2 <= γ, then the measurement is considered to be γ. Falling into the target Within the gate, valid association pairs are formed; otherwise, they are discarded. A validation matrix is ​​generated, with elements corresponding to valid association pairs being 1, and others being 0.

[0102] 3. Calculate the initial association weights: For each valid association pair ( , ), calculate its initial association likelihood weight β_{ }。 β_{ It can be proportional to the measurement z_ (k) The probability density function value at the predicted location of target i. Simultaneously, calculate the weight β_{ for missed detection of the target. 0}.

[0103] 4. Construct a bipartite factorization plot: (e.g.) Figure 2 As shown, a bipartite factorial graph is constructed based on the validation matrix. The left-hand nodes represent targets (first-type nodes), and the right-hand nodes represent measurements (second-type nodes). Only when the validation matrix contains ( , Only when the element is 1 is it in the target node. and measurement nodes Establish an edge between them. Initialize the messages passed on all edges (e.g., initialize them all to 1).

[0104] 5. Message Iteration: Perform multiple rounds of message passing iterations on the constructed bipartite factorization graph until the convergence condition is met. The iteration process is as follows: Figure 3 As shown, it specifically includes: a. Calculate target -> measurement message (corresponding to the first message calculation step): For each edge ( , ), calculate from the target Send to Measurement intermediate messages (Round p). The calculation formula can be: ,in This is the scaling factor.

[0105] b. Update target -> measurement message (corresponding to the first message update step): Apply damping factor Smooth the calculated intermediate messages: This step corresponds to "applying a damping factor to the outgoing message for weighted smoothing".

[0106] c. Calculate the measurement -> target message (corresponding to the second message calculation step): For each edge ( , ), calculation from measurement Send to target intermediate messages The calculation formula reflects the constraint that "a measurement originates from at most one target" and can introduce a scaling factor: .

[0107] d. Update Measurement -> Target Message (corresponding to the second message update step): Apply the damping factor as well. Smooth: .

[0108] e. Convergence Check: Calculate the maximum absolute difference Δ between all messages μ and ν in the current round and their corresponding messages in the previous round. If Δ < ε (ε is a preset convergence threshold, such as 1e-5) or the number of iterations p reaches the maximum number of iterations P_max (such as 50), then stop the iteration; otherwise, p = p + 1, and return to step a to continue the iteration.

[0109] 6. Calculate marginal probabilities and update the state: a. Calculate the marginal association probability: After iterative convergence, for each valid association pair ( , ), calculate its edge association probability. The calculation formula incorporates a scaling factor: And normalize it so that each target The sum of all associated probabilities (including the probability of missed detection) is 1.

[0110] b. State update: using the calculated Update each target according to the standard procedure of the Joint Probabilistic Data Association (JPDA) algorithm. State estimation _ (k|k) and covariance P_ (k|k). The update process includes the calculation of combined innovation and covariance spread terms.

[0111] Component name, connection / location relationship: Processor: Executes the above algorithm steps.

[0112] Memory: Stores program code, target status, measurement data, intermediate messages, probability values, etc.

[0113] Sensor interface: Receives measurement data input from external sensors (such as radar).

[0114] The steps are connected in the logical order described above, and the data flow is from the output of the previous step to the input of the next step.

[0115] Working principle: The working principle of this invention is based on graph model reasoning. By constructing a bipartite factor graph, the complex global data association problem is transformed into a message passing problem between local nodes. A damping factor is introduced. The increased "inertia" of the iteration process smooths message updates, prevents oscillations, and promotes convergence. A scaling factor is introduced. This reduces the message over-amplification effect ("echo") caused by loops in the graph, corrects the calculated probability distribution to be closer to the true posterior distribution, and thus avoids "overconfidence." Finally, by using more accurate probabilities for state updates, a more precise and robust target state estimate is obtained.

[0116] Dimensions: Damping factor The value range is [0, 1), with a typical value of 0.5 ± 0.2. In complex tracking situations, such as when the target position is close, the value should be appropriately reduced.

[0117] scaling factor Value range (0, 1], typical value 0.8±0.1. In complex tracking situations, such as when the target position is close, the value should be appropriately reduced.

[0118] Gate threshold γ: Determined based on the measurement dimension and expected confidence level. For two-dimensional measurements, a typical value is 9.21 (corresponding to 99% confidence level of the chi-square distribution under two-dimensional measurements).

[0119] Convergence threshold ε: range (0, 1e-4), typical value 1e-6.

[0120] Maximum number of iterations P max The optimal range is 50~100, and 100 is used in the formal experiment.

[0121] Implementation steps: System startup and initialization: Start the signal processing system, load the program, and initialize the target trajectory (initial state and covariance).

[0122] Data Acquisition: At time k, acquire a new measurement dataset Z(k) = {z_ (k)}.

[0123] Execution tracking cycle: for each existing target track The above steps are executed in sequence: 1 (state prediction) -> 2 (gate filtering) -> 3 (calculate initial weights) -> 4 (build factor graph) -> 5 (message iteration) -> 6 (probability calculation and state update).

[0124] Track Management: Based on the updated status and association probability, perform management operations such as track initiation, confirmation, and deletion.

[0125] Output: Output the state estimates of all targets at the current time k.

[0126] Execute in a loop: k = k+1, return to step 2, and process the measurement data at the next time step.

[0127] Effect verification: To verify the effectiveness of the method in this embodiment, the following simulation experiment was designed: Experimental scenario: Simulate a scenario containing multiple moving targets (e.g., 8 targets) whose trajectories intersect. Set a high clutter density (e.g., an average of 70 clutter measurements per scan cycle) and a certain detection probability (e.g., 0.85).

[0128] Comparison method: The method of this embodiment (denoted as CBP-JPDA) is compared with the traditional cyclic belief propagation joint probability data association method without the introduction of damping and scaling factors (denoted as LBP-JPDA).

[0129] Evaluation Metric: The Expected Optimal Submode Assignment (E-OSPA) distance is used as the core metric, with parameters set as follows: order = 2, cutoff distance = 5. The E-OSPA distance can simultaneously measure state estimation error and target quantity estimation error.

[0130] Experimental procedure: Conduct multiple (e.g., 50) Monte Carlo simulation experiments and statistically analyze the average performance.

[0131] Results: Plot the average E-OSPA distance of the two methods as a function of time steps (see the schematic diagram for the results). Figure 4 ).

[0132] The curve of the method in this embodiment (CBP-JPDA) will be below the curve of the traditional method (LBP-JPDA), indicating that its tracking error is smaller. Statistics such as the global average E-OSPA and the second-half average E-OSPA can be recorded for comparison.

[0133] Result Record Table:

[0134] Example 3: This embodiment illustrates the implementation of the above method as a signal processing system and a computer-readable storage medium. A signal processing system for multi-target tracking includes: a memory for storing a computer program; and a processor for executing the computer program in the memory to implement the multi-target tracking method as described in any one of Embodiment 2.

[0135] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the multi-target tracking method as described in any one of Embodiment 2.

[0136] Detailed Technical Solution: The signal processing system can be an industrial computer, an embedded system (such as a DSP, FPGA, ARM processor, etc.), or a server. The memory can be any medium that can be used to store programs, such as flash memory, hard disk, or RAM, and it stores program code that implements the multi-target tracking method as described in Embodiment 1. The processor is coupled to the memory and configured to read and execute the program code in the memory, thereby controlling the system to execute the steps of the method described in Embodiment 1 after receiving sensor data, and outputting the target tracking result.

[0137] The computer-readable storage medium may be a USB flash drive, optical disc, hard disk, or network server storage space, etc. When the computer program / instruction stored thereon is executed by a processor (such as the processor of the aforementioned signal processing system, or the processor of another computer), the processor will perform the aforementioned multi-target tracking method.

[0138] Effect: This embodiment extends the method invention to the product invention, providing a hardware and software carrier for implementing the method, thus broadening the scope of protection and application scenarios.

[0139] Example 4: This example illustrates the application of the above signal processing system to specific fields such as motor vehicles and drones.

[0140] A motor vehicle includes a signal processing system as described in Embodiment 3 for realizing multi-target tracking in environmental perception and autonomous driving. An air traffic control system includes a signal processing system as described in Embodiment 3 for monitoring and tracking the status of aircraft in airspace. An unmanned aerial vehicle (UAV) includes a signal processing system as described in Embodiment 3 for realizing multi-target tracking in autonomous navigation and obstacle avoidance.

[0141] Detailed Technical Solution: In motor vehicles (such as autonomous vehicles), the signal processing system receives data from sensors such as onboard radar, lidar, and cameras, and executes the method of this invention to track the position and speed of surrounding vehicles, pedestrians, obstacles, and other targets in real time, providing a basis for path planning and decision-making control. In air traffic control systems, the signal processing system receives radar data, executes the method of this invention, and accurately tracks the flight paths of various aircraft within the controlled airspace to ensure flight safety. In unmanned aerial vehicles (UAVs), the signal processing system receives data from its own sensors, executes the method of this invention, and achieves perception and tracking of its surrounding environment and other UAVs / obstacles for autonomous navigation and obstacle avoidance.

[0142] This embodiment demonstrates the practical application value of the method of the present invention in several important fields. The high-precision and high-efficiency tracking capabilities it brings can significantly improve the performance and security reliability of these application systems.

[0143] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-target tracking method, characterized in that, include: The state prediction step includes predicting the target state at the current time and the prediction covariance based on the target state estimate at the previous time step; The factor graph construction step includes establishing a bipartite factor graph model, which includes a first type of node representing the target, a second type of node representing the measurement, and edges connecting the first type of node and the second type of node, wherein the edges are used to represent the association probability weights between the target and the measurement; the message iteration step includes performing multiple rounds of message passing between the first type of node and the second type of node on the bipartite factor graph model, wherein in at least one round of message passing, a damping factor is applied to the outgoing message for weighted smoothing, and / or a scaling factor is applied to the incoming message for confidence correction; the probability calculation and state update step includes calculating the edge association probability between the target and the measurement based on the converged result of the message iteration step, and updating the target state estimate and estimated covariance at the current time based on the edge association probability.

2. The method according to claim 1, characterized in that, Before the factor graph construction step, the method further includes a gate screening step, which includes calculating the statistical distance between the measurement and the target predicted value, and screening out effective association pairs according to a preset threshold. The edges in the bipartite factor graph model are established only based on the effective association pairs.

3. The method according to claim 1, characterized in that, The message iteration steps include: a first message calculation step, which includes calculating the message sent from a first type of node to a second type of node based on the initial association weight of the target and measurement and other messages passed in from the second type of node, and introducing the scaling factor in the calculation; a first message update step, which includes weighting and fusing the message obtained from the first message calculation step with the corresponding historical message from the previous iteration using the damping factor to generate the final message sent to the second type of node; a second message calculation step, which includes calculating the message sent from a second type of node to a first type of node based on the initial association weight of the target and measurement and other messages passed in from the first type of node, and introducing the scaling factor in the calculation; and a second message update step, which includes weighting and fusing the message obtained from the second message calculation step with the corresponding historical message from the previous iteration using the damping factor to generate the final message sent to the first type of node.

4. The method according to claim 3, characterized in that, The calculations in the first message calculation step follow the following mathematical relationships: ,in For the target in the p-th iteration Send to Measurement Intermediate message, As the initial association weight, For other measurements in the previous round Send to target The news The scaling factor is mentioned.

5. The method according to claim 3, characterized in that, The calculations in the second message calculation step follow the following mathematical relationships: ,in Measurement in the p-th iteration Send to target Intermediate message, As the initial association weight, For other objectives in the current round Send to Measurement The news, The scaling factor embodies the constraint that "a measurement can originate from at most one target".

6. The method according to claim 1, characterized in that, In the probability calculation and state update steps, the scaling factor is introduced to correct the probability distribution when calculating the edge association probability. Its calculation follows the following mathematical relationship: ,in For the goal With measurement Edge association probability, As the initial association weight, For all measurements after iterative convergence Send to target The news, The scaling factor is mentioned.

7. The method according to claim 1, characterized in that, The state prediction step and the probability calculation and state update step are implemented based on the Kalman filter framework.

8. The method according to claim 1, characterized in that, The damping factor has a value range of [0, 1), and the scaling factor has a value range of (0, 1).

9. The method according to claim 1, characterized in that, The termination condition for the message iteration step is: the maximum difference between the message matrices generated by two consecutive iterations is less than a preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations.

10. A signal processing system for multi-target tracking, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program in the memory to implement the multi-target tracking method as described in any one of claims 1 to 9.

Images