Target automatic detection and tracking system
By integrating multiple data correlation algorithms and dynamic correlation wavegate technology, the real-time and robustness of multi-objective tracking systems in complex scenarios is solved, and a large number of targets is tracked efficiently and stably, which is suitable for military monitoring, smart transportation and spacecraft monitoring and other fields.
Patent Information
- Application Number
- CN202510582315.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-07-25
AI Technical Summary
The existing multi-objective tracking system is insufficient in real-time, accuracy and robustness in complex scenarios, making it difficult to track a large number of targets at the same time, and in high-density target scenarios, the calculation complexity is high, the anti-interference ability is weak, and the dynamic adaptability is insufficient.
Fusion of nearest neighbor method, global nearest neighbor method, probability data association and joint probability data association algorithm, combined with dynamic correlation wave gate and Kalman filter enhancement update, realize adaptive tracking gate generation and efficient data association, and optimize the calculation complexity through dynamic algorithm switching and joint event pruning strategy.
The goal of 2048 stable tracking of a single site has been achieved, the tracking accuracy has been improved by 40%, the false alarm rate has been reduced to below 0.1%, and the computing efficiency has been improved by 60%, adapting to complex environments and improving the real-time and robustness of the system.
Smart Images
Figure CN120370281A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of multi-target tracking technology, and more specifically, to a system for automatic target detection and tracking. Background Art
[0002] In the field of target detection and tracking, traditional multi-target tracking systems face many technical bottlenecks, especially in terms of real-time performance, accuracy, and robustness in complex scenarios. First of all, existing systems are usually limited by the target processing capacity of a single site. Most solutions can only track hundreds of targets simultaneously, making it difficult to meet the requirements of high-density target scenarios such as military surveillance, intelligent transportation, or drone swarm cooperation. For example, in dense urban traffic monitoring, due to limited computing resources, traditional systems often experience target loss or response delay, severely restricting the real-time decision-making ability. Secondly, the contradiction between the efficiency and accuracy of data association algorithms is prominent: simple association strategies based on the nearest neighbor method (NN) have low computational complexity, but are prone to false tracking or missed detection when target trajectories cross or there is clutter interference; while complex algorithms such as joint probabilistic data association (JPDA) can improve anti-interference ability, but due to the need to enumerate all possible joint events, the computational complexity increases exponentially, making it difficult to be actually deployed. In addition, existing systems have insufficient adaptability to dynamic targets. The fixed-size tracking gate cannot effectively match the position uncertainty of high-speed maneuvering targets, often resulting in tracking failure due to prediction deviation. In an environment with dense clutter, differentiating false alarm echoes from real targets also becomes a problem. Traditional probabilistic data association (PDA) lacks a dynamic adjustment mechanism for the competition weights of multiple targets, further exacerbating the state estimation error. Although existing research has tried to alleviate the above problems through algorithm optimization or hardware upgrade, there are still significant shortcomings in ultra-large-scale target tracking, real-time processing of gate overlap scenarios, and anti-interference ability in complex environments.
[0003] Therefore, there is a need for a system for automatic target detection and tracking to break through the existing technical bottlenecks and meet the reliable tracking requirements for large-scale and high-dynamic targets in practical applications. Summary of the Invention
[0004] To solve the above technical problems, this application is proposed. By integrating the nearest neighbor method, global nearest neighbor method, probabilistic data association, and joint probabilistic data association algorithms, combined with dynamically adjusted association gates and Kalman filter enhanced updates, this application realizes the efficient and stable tracking of 2048 targets simultaneously at a single site.
[0005] According to one aspect of the present application, a system for automatic target detection and tracking is provided, which includes: a sensor interface module for receiving and preprocessing the original measurement data of multi-source sensors; a target recognition module configured to extract initial target tracks through a clustering algorithm and assign unique identifiers; a dynamic association gate module for generating an adaptive tracking gate based on the Kalman filter prediction result and screening candidate echoes; a multi-algorithm association module for supporting the dynamic switching of the nearest neighbor method, the global nearest neighbor method, the probabilistic data association, and the joint probabilistic data association; a state estimation module for updating the Kalman filter through a weighted equivalent innovation vector and outputting the target state and confidence; and a track management module for dynamically creating or terminating tracks to optimize resource allocation.
[0006] Preferably, the dynamic association gate module includes: a tracking gate calculation unit for calculating the geometry and size of the tracking gate in real time according to the predicted value of the target motion state and the covariance matrix; a candidate echo screening unit for screening candidate echoes falling within the tracking gate using a statistical distance threshold, where the statistical distance is the Mahalanobis distance or the Euclidean distance; and a target tracking unit for expanding the coverage range of the tracking gate for high-speed maneuvering targets and shrinking the tracking gate for low-speed targets to exclude clutter.
[0007] Preferably, the dynamic switching logic of the multi-algorithm association module includes: when the target density is lower than a preset threshold and there is no gate overlap, enabling the global nearest neighbor method combined with the Hungarian algorithm for fast matching; when the target density is higher than the preset threshold or gate overlap is detected, switching to the joint probabilistic data association and reducing the computational complexity through the confirmation matrix splitting and joint event pruning strategy.
[0008] Preferably, the optimized implementation of the joint probabilistic data association includes: constructing a confirmation matrix representing the coverage relationship between the measurement and the target tracking gate; splitting the confirmation matrix to generate an interconnection matrix and calculating the joint interconnection probability between the measurement and multiple targets; where the splitting principle of the confirmation matrix is: only one non-zero element is retained in each row to ensure the unique source of the measurement; and except for the clutter column, at most one non-zero element is retained in each column to ensure that each target is associated with at most one measurement.
[0009] Preferably, the update method of the state estimation module includes: a prediction unit for generating predicted values of the target position, velocity, and covariance through the Kalman filter based on the historical state; an update unit for calculating the equivalent innovation vector according to the association result, and the formula is:
[0010]
[0011] where, is the equivalent innovation vector for target t at time step k, m k is the number of valid candidate measurements falling within the association gate of target t at time step k, βjt (.) is the joint association probability between the j-th measurement and target t at time step k, z j is the observation value vector of the j-th measurement; an output result unit for outputting the final state estimation of the target and the updated covariance matrix.
[0012] Preferably, the track management module includes: starting an end timer for tracks that are continuously unassociated with measurements, and releasing resources after timeout; assigning an initial covariance matrix to new target tracks and dynamically adjusting the target motion model parameters according to the association success rate.
[0013] According to one aspect of the present application, there is also provided a system method for automatic target detection and tracking, including: S1, receiving sensor measurement data and performing coordinate transformation and noise filtering; S2, extracting potential target points and initializing tracks; S3, generating a dynamic tracking gate based on Kalman filtering prediction to screen candidate echoes; S4, selecting a data association algorithm according to the scene complexity and calculating the matching relationship between measurements and tracks; S5, updating the target state through a weighted equivalent innovation vector and outputting the tracking result; S6, dynamically managing the track life cycle to optimize system resources.
[0014] Preferably, in the step S3, the method for generating the dynamic tracking gate includes: calculating the lengths of the major and minor axes of the tracking gate according to the target speed and covariance matrix; adding a tracking gate size expansion coefficient for maneuvering targets, and the calculation formula is:
[0015] ρ = 1 + α·||a k ||
[0016] where α is a preset parameter, a k is the target acceleration; for stationary or low-speed targets, a rectangular tracking gate with a fixed size is adopted.
[0017] Preferably, in the step S4, the selection logic of the data association algorithm includes: when it is detected that the gate overlaps or the clutter density exceeds the threshold, perform the following operations: S41, constructing a confirmation matrix and splitting it to generate a set of association events; S42, calculating the joint probability of each event and pruning events with a probability lower than 1%; S43, updating the track state according to the weights of the remaining events; when there is no gate overlap and the clutter density is lower than the threshold, directly solve the global optimal matching through the Hungarian algorithm.
[0018] According to one aspect of the present application, there is also provided a computer-readable storage medium, characterized in that it stores program instructions, and when the instructions are executed by a processor, the multi-target tracking method described in any one of the above methods is implemented.
[0019] The system for automatic target detection and tracking of the present invention includes a dynamic algorithm switching mechanism (using GNN to optimize efficiency in low clutter scenarios and enabling JPDA to improve accuracy in high clutter scenarios), intelligent gate generation based on the target motion model (adaptive shape and size), a joint event pruning strategy (reducing the computational complexity of JPDA), and weighted equivalent innovation vector update (resisting occlusion and cross-interference). This technology significantly solves problems such as traditional systems' handling of ultra-large-scale targets, mis-association in complex scenarios, low computational efficiency, and insufficient dynamic adaptability. It has been successfully applied in fields such as military surveillance, intelligent transportation, and spacecraft monitoring, achieving a 40% improvement in tracking accuracy and reducing the false alarm rate to less than 0.1%, with both high real-time performance and strong robustness.
[0020] The beneficial effects of the present invention are as follows:
[0021] Ultra-large-scale target handling ability: It supports stable tracking of 2048 targets simultaneously at a single site, breaking through the capacity limit of traditional systems (usually hundreds), and meeting the requirements of high-density scenarios such as military surveillance and intelligent transportation.
[0022] Significantly improved real-time performance: Through the dynamic algorithm switching mechanism (using GNN + Hungarian algorithm in low clutter and JPDA in high clutter), the computational efficiency is increased by more than 60%; the joint event pruning strategy reduces the computational complexity of JPDA from exponential to linear, solving the real-time processing problem in the scenario of gate overlap.
[0023] Enhanced anti-interference and tracking accuracy: In scenarios with dense clutter or target crossing, the mis-association rate is reduced by 30% - 50%; through weighted equivalent innovation vector update and probabilistic data association (PDA / JPDA), the state estimation accuracy is significantly improved when there is occlusion or trajectory crossing, and the missed detection rate is less than 5%.
[0024] Optimized dynamic adaptability: The adaptive tracking gate based on the target motion model (such as Kalman filter) adjusts the shape and size of the gate in real time, and the tracking stability of high-speed maneuvering targets is increased by 40%; it supports flexible switching between multiple scenarios (such as vehicles, pedestrians, spacecraft), meeting the requirements of complex environments. Description of the Drawings
[0025] By describing the embodiments of the present application in more detail in combination with the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.
[0026] Figure 1 It is a framework diagram of the system for automatic target detection and tracking according to the embodiments of the present application.
[0027] Figure 2 It is a flowchart of a method for automatic target detection and tracking according to an embodiment of the present application. Detailed implementation manners
[0028] Various exemplary embodiments, features, and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.
[0029] The special term "exemplary" here means "serving as an example, an embodiment, or illustrative". Any embodiment described as "exemplary" here does not have to be construed as superior to or better than other embodiments.
[0030] In addition, for better illustration of the present application, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present application can be implemented without some specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present application.
[0031] Furthermore, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, "a plurality of" means two or more unless otherwise specifically defined.
[0032] Although today's technologies have improved the efficiency of anti-money laundering work to a certain extent, they still cannot achieve full-process intelligence from data processing to report generation. Existing systems lack reasonable task decomposition and tool invocation capabilities, resulting in difficulties in anti-money laundering work and making it difficult to meet the requirements of modern financial institutions for the real-time and accuracy of anti-money laundering work.
[0033] The present invention proposes a system for automatic target detection and tracking. The system includes a sensor interface module, a target recognition module, a dynamic association gate module, a multi-algorithm association engine, a state estimation module, and a track management module. By fusing the nearest neighbor method (NN), the global nearest neighbor method (GNN), the probabilistic data association (PDA), and the joint probabilistic data association (JPDA) algorithms, combined with dynamic tracking gate screening and Kalman filter enhanced update, it can achieve stable tracking of 2048 targets simultaneously at a single site. This technology significantly improves the tracking real-time performance, anti-interference ability, and accuracy in complex scenarios, and is applicable to fields such as military surveillance, intelligent transportation, unmanned aerial vehicle clusters, and spacecraft monitoring.
[0034] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods:
[0035] Figure 1 It is a framework diagram of a system for automatic target detection and tracking according to an embodiment of the present application. As Figure 1 shown, the automatic target detection and tracking system 100 according to the embodiment of the present application includes: a sensor interface module 101 for receiving and preprocessing the raw measurement data of multi-source sensors; a target recognition module 102 configured to extract initial target tracks through a clustering algorithm and assign unique identifiers; a dynamic association gate module 103 for generating an adaptive tracking gate based on the Kalman filter prediction result and screening candidate echoes; a multi-algorithm association module 104 for supporting the dynamic switching of the nearest neighbor method, the global nearest neighbor method, the probabilistic data association, and the joint probabilistic data association; a state estimation module 105 for updating the Kalman filter through a weighted equivalent innovation vector and outputting the target state and confidence; and a track management module 106 for dynamically creating or terminating tracks to optimize resource allocation.
[0036] In the embodiment of the present application, the sensor interface module 101 is used to receive and preprocess the raw measurement data of multi-source sensors. It should be understood that the data formats output by devices such as radars, cameras, and lidars are diverse (such as the range-velocity information of radars, the image frames of cameras, and the point clouds of lidars), and need to be uniformly converted into a standardized format that can be parsed by the system through the interface module. Single sensors have limitations (such as radars being unable to recognize colors and cameras being affected by lighting), and multi-source data fusion can improve the comprehensiveness and robustness of target detection. Sensor data often contains environmental noise (such as radar multipath effects and camera image blur) and hardware errors (such as lidar ranging drift). The preprocessing operations include: using Kalman filtering or median filtering to eliminate random noise; converting polar coordinates (radars) to Cartesian coordinates for convenient unified processing; normalizing the numerical ranges of different sensors (such as pixel intensities and distance values) to prevent algorithms from failing due to dimensional differences. By integrating multi-source data, suppressing noise, and unifying the spatio-temporal reference, high-quality input is provided for subsequent target recognition and tracking, ensuring the high precision, high reliability, and scalability of the system in complex scenarios.
[0037] In the embodiment of the present application, the target recognition module 102 is configured to extract the initial target track through a clustering algorithm and assign a unique identifier. It should be understood that the preprocessed data output by multi-source sensors (such as radar, camera) is usually a discrete set of point traces or pixels, including noise, clutter, and multi-target mixed information. Point traces that are spatially or feature-wise close are grouped into the same cluster, and each cluster represents a potential target. For example, an area where radar point traces are densely distributed in space may correspond to a vehicle or a drone. By noise filtering, isolated points (such as sensor noise or transient interference) are removed, and valid target point traces are retained. The clustering algorithms that can be selected include DBSCAN (density clustering), K-means, and Mean-Shift. DBSCAN (density clustering) is suitable for processing data with a lot of noise and irregular cluster shapes (such as point traces of a vehicle in complex road conditions). K-means is applicable to scenarios where the cluster shapes are regular and the target distribution is uniform (such as a group of drones flying in formation in the air). Mean-Shift dynamically adapts to changes in target density and is suitable for real-time clustering of moving targets (such as vehicles in traffic flow). Through clustering, the original point traces are compressed into structured targets, reducing the computational complexity of subsequent association algorithms. Filtering noise point traces avoids false alarm targets from interfering with the tracking process. An initial motion model (such as Kalman filter) is established for each target to support subsequent state prediction and update. The unique ID ensures that the target can still be correctly identified when it is occluded, crossed, or reappears (such as a vehicle reappearing after briefly driving into a tunnel).
[0038] In the embodiment of the present application, the dynamic association gate module 103 generates an adaptive tracking gate based on the Kalman filter prediction result and screens candidate echoes. It should be understood that although the clustering algorithm can extract the initial target position and assign an ID, the target moves over time and the sensor data is continuously updated. It is necessary to dynamically predict the future state of the target to maintain the continuity of tracking. Centered on the Kalman filter prediction value, a region (such as an ellipse, rectangle) is delimited according to the covariance matrix, and only the candidate echoes within this region are considered for association.
[0039] Specifically, in a specific example of the present application, the dynamic association gate module includes: a tracking gate calculation unit for calculating the geometry and size of the tracking gate in real time according to the predicted value of the target motion state and the covariance matrix; a candidate echo screening unit for screening candidate echoes falling within the tracking gate using a statistical distance threshold, where the statistical distance is the Mahalanobis distance or the Euclidean distance; and a target tracking unit for expanding the coverage range of the tracking gate for high-speed maneuvering targets and shrinking the tracking gate for low-speed targets to exclude clutter.
[0040] In the embodiment of the present application, the multi-algorithm association module 104 is used to support the dynamic switching of the nearest neighbor method, the global nearest neighbor method, probabilistic data association, and joint probabilistic data association. It should be understood that first, for the nearest neighbor method (NN), the working principle of the nearest neighbor standard filter is to first set a tracking gate, and the echoes obtained by the preliminary screening of the tracking gate (correlation wave gate) become candidate echoes to limit the number of targets participating in the correlation discrimination. Assume that before the k-th scan, N tracks have been established. The new measurement at the k-th time is z j (k), j = 1, 2, …, N. Inside the association gate of the i-th track, the difference vector between the observation j and the track i is defined as the difference between the measurement value and the predicted value, that is, the filter residual is
[0041]
[0042] Let S(k) be the covariance matrix of e ij (k), then the square of the statistical distance is
[0043] d ij = e ij (k) s(k) -1 e ij T (k)
[0044] The echo with the distance value within the threshold range and the closest distance can be considered successfully associated. The advantage of the nearest neighbor method is its simple calculation. The disadvantage is that in a multi-echo environment, the candidate echo closest to the predicted position of the target (especially targets that are relatively close or have intersecting trajectories) is not necessarily the true wave of the target. That is, the correlation performance of this method is imperfect, and there may be phenomena of false tracking and target loss.
[0045] The distance calculation method of the global nearest neighbor method (GNN) is the same as that of NN, but it is an optimal allocation problem to minimize the total distance or the association cost.
[0046]
[0047] When the association matrix is large, for the two-dimensional matching problem, the KM algorithm or the Hungarian algorithm can be used to solve it.
[0048] The probabilistic data association algorithm (PDA) believes that all valid measurements that appear within the association wave gate may originate from the target, but the probability of each valid measurement originating from the target is different. This algorithm calculates and gives the probability weight of each valid measurement, and uses the weighted value of the valid measurement as the final measurement estimate value, which is introduced into the update step of the Kalman filter.
[0049] Prediction (the same as the standard KF)
[0050]
[0051] Measurement prediction (calculate the innovation vector for all valid measurements within the associated gate)
[0052]
[0053] Update (weighted sum of the innovation vectors corresponding to all measurement predictions to obtain the equivalent innovation vector)
[0054]
[0055] where, β i,k represents the probability that the i-th measurement at time k originates from this target under the condition of knowing all valid observations. This probability is calculated according to the following formula
[0056]
[0057] The Joint Probabilistic Data Association (JPDA) algorithm is proposed based on the Probabilistic Data Association (PDA) algorithm which is only applicable to single-target tracking. This method is a good algorithm for data association of multiple targets in a clutter environment. Its principle is similar to that of PDA. JPDA also calculates a weighted residual for track update based on all measurements within the validation gate. The difference is that when an echo falls into the overlapping area of the correlation gates of different targets, it is necessary to comprehensively consider the source of each measurement. When calculating the association probability, the competition of multiple tracks for the measurement needs to be considered, and the weighted value of the competition should be reduced to reflect the competition of other targets for this measurement. The multi-target data association technology in a clutter environment is the most important and difficult problem in multi-target tracking. If the correlation gates of multiple targets being tracked do not intersect, or no echo falls into the intersection area of the gates, the multi-target data association problem can be simplified into multiple single-target data association problems, and the Probabilistic Data Association algorithm can be used to solve it. However, if an echo falls into the intersection area of the correlation gates of each target, the data association problem at this time will be much more complex.
[0058] When an echo falls into the overlapping area of the correlation gates of different targets, it is necessary to comprehensively consider the source of each measurement. To represent the complex relationship between valid echoes and the tracking gates of each target, the concept of a validation matrix is introduced. The validation matrix is defined as
[0059]
[0060] In the formula, ω jt is a binary variable, ω jt represents that the measurement j (j = 1, 2,..., m k ) falls into the validation gate of target t (t = 0, 1,..., T), and ω jt= 0 means that the measurement does not fall within the confirmation gate of target t. t = 0 means there is no target, and at this time, the column elements ω of Ω j0 are all 1, because each measurement may originate from clutter or be a false alarm.
[0061] Specifically, in a specific example of the present application, the dynamic switching logic of the multi-algorithm association module includes: when the target density is lower than the preset threshold and there is no gate overlap, enabling the global nearest neighbor method combined with the Hungarian algorithm for fast matching; when the target density is higher than the preset threshold or gate overlap is detected, switching to joint probabilistic data association, and reducing the computational complexity through the confirmation matrix splitting and joint event pruning strategy.
[0062] Further, the optimized implementation of the joint probabilistic data association includes: constructing a confirmation matrix to represent the coverage relationship between the measurement and the target tracking gate; splitting the confirmation matrix to generate an interconnection matrix and calculating the joint interconnection probability of the measurement and multiple targets; wherein, the splitting principle of the confirmation matrix is: only one non-zero element is retained in each row to ensure the unique source of the measurement; except for the clutter column, at most one non-zero element is retained in each column to ensure that each target is associated with at most one measurement.
[0063] For a given multi-target tracking problem, once the confirmation matrix (or interconnection aggregation matrix) Ω reflecting the interconnection situation between the valid echo and the target or clutter is given, all the interconnection matrices representing the interconnection events can be obtained by splitting the confirmation matrix. When splitting the confirmation matrix, two basic assumptions must be followed:
[0064] a) Each measurement has a unique source, that is, if any measurement does not originate from a certain target, it must originate from clutter, that is, the measurement is a false alarm. In other words, the situation of indistinguishable detections is not considered here.
[0065] b) For a given target, at most one measurement has it as the source. If a target may be matched with multiple measurements, one will be taken as true and the others as false.
[0066] That is to say, two principles must be followed when splitting the confirmation matrix.
[0067] a) In each row of the confirmation matrix, select one and only one 1 as the only non-zero element of the interconnection matrix in that row. This is actually to make the possible joint events represented by the feasible matrix satisfy the first assumption, that is, each measurement has a unique source.
[0068] b) In the feasible matrix, except for the first column, at most one non-zero element can be retained in each column. This is to make the feasible events represented by the interconnection matrix satisfy the second assumption, that is, each target has at most one measurement with it as the source.
[0069] The purpose of joint probabilistic data association is to calculate the probability of association between each measurement and its possible source targets. When there are echoes falling into the overlapping area of the correlation gates of different targets, the source situation of each measurement must be comprehensively considered. Let θ j,t (k) denote the event that measurement j originates from target t (0 ≤ t ≤ T), and the event θ j0 (k) denote that measurement j originates from clutter or false alarm. According to the definition of conditional probability in the single-target probabilistic data association filter, we have
[0070] β jt (k) = Pr{θ jt (k)|Z k}, j = 0, 1, …, m k , t = 0, 1, …, T
[0071] denote the probability of association between the j-th measurement and target t, and
[0072]
[0073] Then the state estimate of target t at time k is
[0074]
[0075] where
[0076]
[0077] denote the state estimate obtained by performing Kalman filtering on target t using the j-th measurement at time k. And denote the situation where no measurement originates from the target at time k. In this case, the predicted value is used instead. The probability of association between the j-th measurement and the target can be obtained using the following formula:
[0078]
[0079] where, denote the time when measurement j originates from target t (0 ≤ t ≤ T) in the i-th joint time, θ i (k) denote the i-th joint event, n k denote the number of joint events, and
[0080]
[0081] denote whether measurement j originates from target t in the i-th joint event. It is 1 when measurement j originates from target t, and 0 otherwise. Define the i-th joint event in the general case as
[0082]
[0083] It represents m k a possibility of the matching of m measurements with different targets. The interconnection matrix corresponding to the joint event is defined as
[0084]
[0085] Based on the above two basic assumptions, it is easy to deduce that the interconnection matrix satisfies:
[0086]
[0087] In the embodiment of the present application, the state estimation module 105 is used to update the Kalman filter through a weighted equivalent innovation vector and output the target state and confidence. It should be understood that in a multi-target tracking scenario, sensor measurements (such as radar echoes, camera images) often contain multiple candidate observations, and these observations may belong to real targets, clutter, or false alarms. Traditional Kalman filtering relies on a single observation for state update, but there are the following problems in a complex environment: if the wrong observation is selected, it will lead to state estimation deviation or even tracking loss; ignoring other potentially valid observations will reduce the update accuracy. The role of the weighted equivalent innovation vector includes generating a comprehensive equivalent measurement by fusing the probability weights of all candidate observations, avoiding relying on a single observation, and enhancing the anti-interference ability.
[0088] Specifically, in a specific example of the present application, the update method of the state estimation module includes: a prediction unit, which is used to generate predicted values of the target position, velocity, and covariance through Kalman filtering based on the historical state; an update unit, which is used to calculate the equivalent innovation vector according to the association result, and the formula is:
[0089]
[0090] where is the equivalent innovation vector for target t at time step k, m k is the number of valid candidate measurements falling within the association gate of target t at time step k, β jt (k) is the joint interconnection probability between the jth measurement and target t at time step k, z j is the observation value vector of the jth measurement; an output result unit, which is used to output the final state estimation of the target and the updated covariance matrix.
[0091] In the embodiment of the present application, the track management module 106 is used to dynamically create or terminate tracks to optimize resource allocation. A multi-target tracking system needs to process a large amount of sensor data (such as radar, camera), and track thousands of targets at the same time. The computing resources (CPU, memory) are limited. If all tracks (including lost or invalid targets) are continuously maintained, it will lead to waste of resources and affect real-time performance; new targets may not be tracked in time due to insufficient resources, resulting in missed detections.
[0092] Specifically, in an embodiment of the present application, the track management module includes: starting an end timer for tracks that are continuously unassociated with measurements, and releasing resources after timeout; assigning an initial covariance matrix to new target tracks, and dynamically adjusting target motion model parameters according to the association success rate.
[0093] Furthermore, Figure 2 is a flowchart of a method for automatic target detection and tracking according to an embodiment of the present application. As Figure 2 shown, the method for automatic target detection and tracking according to an embodiment of the present application includes the steps of: S1, receiving sensor measurement data and performing coordinate transformation and noise filtering; S2, extracting potential target traces and initializing tracks; S3, generating a dynamic tracking gate based on Kalman filter prediction, and screening candidate echoes; S4, selecting a data association algorithm according to the scene complexity, and calculating the matching relationship between measurements and tracks; S5, updating the target state through a weighted equivalent innovation vector, and outputting a tracking result; S6, dynamically managing the track life cycle to optimize system resources.
[0094] Correspondingly, in the step S3, the method for generating a dynamic tracking gate includes: calculating the lengths of the major axis and minor axis of the tracking gate according to the target speed and covariance matrix; adding a tracking gate size expansion coefficient to maneuvering targets, and the calculation formula is:
[0095] ρ = 1 + α·||a k ||
[0096] where α is a preset parameter, and a k is the target acceleration; a fixed-size rectangular tracking gate is used for stationary or low-speed targets.
[0097] Specifically, in the step S4, the selection logic of the data association algorithm includes: when it is detected that the gate overlaps or the clutter density exceeds the threshold, perform the following operations: S41, constructing a confirmation matrix and splitting it to generate a set of association events; S42, calculating the joint probability of each event, and pruning events with a probability lower than 1%; S43, updating the track state according to the weights of the remaining events; when there is no gate overlap and the clutter density is lower than the threshold, directly solve the global optimal matching through the Hungarian algorithm.
[0098] In summary, by integrating the Nearest Neighbor (NN), Global Nearest Neighbor (GNN), Probabilistic Data Association (PDA), and Joint Probabilistic Data Association (JPDA) algorithms, combined with dynamically adjusted association gates and enhanced Kalman filter updates, efficient and stable tracking of 2048 targets is achieved simultaneously at a single site. The core of the system includes a dynamic algorithm switching mechanism (optimizing efficiency with GNN in low clutter scenarios and enabling JPDA to improve accuracy in high clutter scenarios), intelligent gate generation based on the target motion model (adaptive shape and size), joint event pruning strategy (reducing the computational complexity of JPDA), and weighted equivalent innovation vector update (resisting occlusion and cross-interference). This technology significantly addresses the problems of traditional systems in processing ultra-large-scale targets, mis-association in complex scenarios, low computational efficiency, and insufficient dynamic adaptability. It has been successfully applied in fields such as military surveillance, intelligent transportation, and spacecraft monitoring, achieving a 40% improvement in tracking accuracy and reducing the false alarm rate to less than 0.1%, with both high real-time performance and strong robustness.
[0099] The embodiments of the present disclosure have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein.
Claims
1. A system for automatic target detection and tracking, characterized in that, Comprising: A sensor interface module, configured to receive raw measurement data from multi-source sensors and perform preprocessing; A target recognition module, configured to extract initial target tracks through a clustering algorithm and assign unique identifiers; A dynamic association gate module, which generates an adaptive tracking gate based on the Kalman filter prediction result to screen candidate echoes; A multi-algorithm association module, which supports dynamic switching between the nearest neighbor method, the global nearest neighbor method, probabilistic data association, and joint probabilistic data association; A state estimation module, which updates the Kalman filter through a weighted equivalent innovation vector and outputs the target state and confidence level; A track management module, which dynamically creates or terminates tracks to optimize resource allocation.
2. The system for automatic target detection and tracking according to claim 1, wherein The dynamic association gate module includes: A tracking gate calculation unit, which calculates the geometry and size of the tracking gate in real time according to the predicted value of the target motion state and the covariance matrix; A candidate echo screening unit, which screens candidate echoes falling within the tracking gate by using a statistical distance threshold, and the statistical distance is the Mahalanobis distance or the Euclidean distance; A target tracking unit, which expands the coverage range of the tracking gate for high-speed maneuvering targets and shrinks the tracking gate for low-speed targets to exclude clutter.
3. The system for automatically detecting and tracking a target according to claim 1, characterized in that, The dynamic switching logic of the multi-algorithm association module includes: When the target density is lower than a preset threshold and there is no gate overlap, the global nearest neighbor method combined with the Hungarian algorithm is enabled for fast matching; When the target density is higher than a preset threshold or gate overlap is detected, it switches to joint probabilistic data association, and the computational complexity is reduced through the confirmation matrix splitting and joint event pruning strategies.
4. The system for automatic target detection and tracking according to claim 3, characterized in that, The optimized implementation of the joint probabilistic data association includes: Constructing a confirmation matrix to represent the coverage relationship between the measurement and the target tracking gate; Splitting the confirmation matrix to generate an interconnection matrix and calculating the joint interconnection probability between the measurement and multiple targets; Among them, the splitting principle of the confirmation matrix is: Only one non-zero element is retained in each row to ensure the unique source of the measurement; Except for the clutter column, at most one non-zero element is retained in each column to ensure that each target is associated with at most one measurement.
5. The system for automatically detecting and tracking a target according to claim 1, wherein, The update method of the state estimation module includes: A prediction unit, which generates predicted values of the target position, velocity, and covariance through the Kalman filter based on the historical state; An update unit, which calculates the equivalent innovation vector according to the association result, and the formula is: Among them, is the equivalent innovation vector for target t at time step k, m k is the number of valid candidate measurements falling within the gate associated with target t at time step k, β jt (k) is the joint association probability between the j-th measurement and target t at time step k, z j is the observation value vector of the j-th measurement; An output result unit, which outputs the final target state estimation and the updated covariance matrix.
6. The system for automatically detecting and tracking a target according to claim 1, wherein The track management module includes: Starting a termination timer for tracks that are continuously not associated with measurements, and releasing resources after timeout; Assigning an initial covariance matrix to new target tracks and dynamically adjusting the target motion model parameters according to the association success rate.
7. A method for automatic target detection and tracking, characterized in that, Comprising: S1, receiving sensor measurement data and performing coordinate transformation and noise filtering; S2, extracting potential target points and initializing tracks; S3, generating a dynamic tracking gate based on the Kalman filter prediction to screen candidate echoes; S4, selecting a data association algorithm according to the scene complexity and calculating the matching relationship between the measurement and the track; S5, updating the target state through a weighted equivalent innovation vector and outputting the tracking result; S6, dynamically managing the track life cycle to optimize system resources.
8. The method for automatically detecting and tracking a target according to claim 7, wherein In the step S3, the method for generating a dynamic tracking gate includes: Calculate the major and minor axis lengths of the tracking gate based on the target speed and covariance matrix; Add a tracking gate size expansion coefficient to the maneuvering target, and the calculation formula is: ρ = 1 + α·||a k || where α is a preset parameter, and a k is the target acceleration; Use a rectangular tracking gate with a fixed size for stationary or low-speed targets.
9. The method for automatically detecting and tracking a target according to claim 7, wherein, In the step S4, the selection logic of the data association algorithm includes: When it is detected that the gate overlaps or the clutter density exceeds the threshold, perform the following operations: S41, construct a confirmation matrix and split it to generate an interconnected event set; S42, calculate the joint probability of each event and prune the events with a probability lower than 1%; S43, update the track state according to the weights of the remaining events; When there is no gate overlap and the clutter density is lower than the threshold, directly solve the global optimal matching through the Hungarian algorithm.
10. A computer-readable storage medium, characterized in that, Store program instructions, and when the instructions are executed by a processor, implement the method for automatic target detection and tracking described in any one of claims 7-9.