A multi-target tracking method based on Gaussian mixture probability hypothesis density filter

Abstract

The application discloses a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering, and belongs to the technical field of multi-target tracking. In order to solve the problems of target missing tracking and track false deletion in the high-density and close multi-target scene, the application obtains the measurement data corresponding to each sensor, and assigns the measurement identifier corresponding to the measurement data; the Gaussian component of each sensor is updated, and the measurement source label of the Gaussian component is set as the measurement identifier corresponding to the sensor measurement data; an adaptive pruning and merging decision model is established; all Gaussian components are input into the adaptive pruning and merging decision model for merging or pruning processing to obtain the remaining Gaussian components; the state extraction weight threshold is set, the remaining Gaussian components are compared, and the mean value of the Gaussian component with the weight higher than the state extraction weight threshold is output as the final extracted target state. The application realizes the significant improvement of the multi-target tracking precision, robustness and continuity.

Description

Technical Field

[0001] This invention belongs to the field of multi-target tracking technology, specifically relating to a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering. Background Technology

[0002] The task of a multi-target tracking system consists of three main stages: a perception stage, which uses sensors such as radar, lidar, or image sensors to scan the observation space and acquire measurement information; a filtering stage, which processes the measurement data returned by the sensors, containing the target's true signal and a large amount of clutter, to achieve target state estimation and trajectory management; and an application stage, which outputs stable and accurate tracking results to the decision-making system for path planning or threat assessment. Tracking performance and accuracy are directly determined by the multi-target tracking filter, and the hypothesis management strategy in the filtering algorithm, namely pruning and merging methods, is a key technology for achieving accurate tracking. Choosing an appropriate hypothesis management strategy is extremely important for the system to accurately perceive the situation.

[0003] Currently, the commonly used hypothesis management strategy in GM-PHD filters is pruning and merging based on fixed thresholds. This strategy is derived and designed under idealized scenarios where targets are sparse and far apart, and under ideal conditions, this method can achieve relatively good tracking results. However, in real-world scenarios, considering the high-density, near-proximity multi-target characteristics in applications such as autonomous driving and drone swarms, the tracking accuracy of this strategy is severely affected: the spatial distance-based merging strategy may incorrectly merge hypotheses representing different real targets, leading to incorrect target number estimation and track loss; while the fixed-weight-based pruning strategy may prematurely delete real target hypotheses with temporarily low weights due to signal fragmentation, thus affecting the overall accuracy and robustness of multi-target tracking. Summary of the Invention

[0004] The problem this invention aims to solve is the issue of target mistracking and track deletion in high-density, near-multi-target scenarios. It proposes a multi-target tracking method based on Gaussian mixture probability assumption density filtering.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-target tracking method based on Gaussian mixture probability hypothesis density filtering includes the following steps:

[0007] S1. At each filtering time, acquire the measurement data corresponding to each sensor and assign the measurement identifier corresponding to the measurement data;

[0008] S2. Based on the measurement data corresponding to each sensor obtained in step S1, update the Gaussian component of each sensor and set the measurement source label of the Gaussian component to the measurement identifier corresponding to the measurement data of that sensor.

[0009] S3. Establish an adaptive pruning and merging decision model;

[0010] S4. Input all Gaussian components into the adaptive pruning and merging decision model established in step S3, and perform merging or pruning to obtain the remaining Gaussian components.

[0011] S5. Set the state extraction weight threshold, compare the retained Gaussian components obtained in step S4, and output the mean of the Gaussian components with weights higher than the state extraction weight threshold as the final extracted target state.

[0012] Furthermore, in step S2, the i-th Gaussian component at time k includes the weight of the Gaussian component, the mean of the Gaussian component, the covariance of the Gaussian component, the measurement source label carried by the Gaussian component, and the timestamp when the measurement source label is assigned.

[0013] Furthermore, the specific implementation method of step S3 includes the following steps:

[0014] S3.1. Traverse all Gaussian component pairs and calculate the state correlation of all Gaussian component pairs using the Bach distance; if the Bach distance is less than the merging distance threshold, proceed to the merging decision process; if the Bach distance is greater than the merging distance threshold, suppress merging and proceed to the next step.

[0015] The merging decision process is set to check the measurement source labels of Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are different, the merging process is suppressed and the next step is not performed. If the measurement source labels of the Gaussian component pairs are the same, the merging process is performed to generate merged Gaussian components.

[0016] S3.2. Perform pruning decisions on all merged Gaussian components and unmerged Gaussian components. If the weight of a Gaussian component is greater than the pruning weight threshold, no pruning is performed. If the weight of a Gaussian component is less than the pruning weight threshold, check the age and protection status of the measurement source labels of the Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are old time labels and are not within the protection period, pruning is performed. If the measurement source labels of the Gaussian component pairs are new time labels, no pruning is performed.

[0017] Furthermore, in step S3.1, the i-th Gaussian component is calculated using the Bach distance. and the j-th Gaussian component The state correlation between them is expressed as:

[0018]

[0019] in, Let the Bach distance be the distance between the i-th Gaussian component and the j-th Gaussian component. and Let be the mean and covariance of the i-th and j-th Gaussian components, respectively. denoted as the average covariance of the i-th Gaussian component and the j-th Gaussian component; det is the value of the matrix determinant.

[0020] Furthermore, in step S3.1, the state correlation is combined with the measurement source label, and a merging decision function based on label dependence is used for judgment. Then, a merging operation is performed on two or more Gaussian components, or the merging operation on two or more Gaussian components is prohibited. The judgment formula is as follows:

[0021]

[0022] in, Let be the decision function for merging the i-th Gaussian component and the j-th Gaussian component. These are the measurement source labels for the i-th Gaussian component and the j-th Gaussian component, respectively. This is the merging threshold based on Bach distance.

[0023] Furthermore, in step S3.2, a pruning decision is made using a dynamic pruning function P based on the labeled age. P outputs 1 to represent retention and 0 to represent pruning. The expression is:

[0024]

[0025] Where u is the unit step function, For component weights, The basic pruning threshold is k, where k is the time step. The labeled age for the component. This is the weight protection period.

[0026] Furthermore, the Gaussian components in the PHD are traversed; it is determined whether the weight of the Gaussian component is lower than a preset pruning weight threshold, and whether the measurement source label of the Gaussian component is a newly assigned label at the current filtering time; if the weight of the Gaussian component is lower than the pruning weight threshold, but its measurement source label is a newly assigned label at the current filtering time, then a weight protection operation is performed on the Gaussian component to prevent components representing newly emerging or reconfirmed real targets from being prematurely deleted.

[0027] Furthermore, the weight protection operation exempts the Gaussian component from pruning during the current filtering cycle; and applies a specific pruning weight threshold lower than the normal pruning weight threshold to the Gaussian component during the current filtering cycle.

[0028] Furthermore, in the prediction step of the GM-PHD filter, the Gaussian component records and retains its measurement source label at all times during the guard period.

[0029] The beneficial effects of this invention are:

[0030] This invention presents a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering, establishing a two-dimensional decision-making system combining "source information" and "state information," fundamentally improving the ability to distinguish nearby targets. Existing technologies, when performing pruning and merging, rely almost entirely on the state information (such as position and velocity) of Gaussian components and their geometric distance in the state space. This single-dimensional criterion has inherent limitations when dealing with multiple nearby targets, failing to distinguish between "two neighboring hypotheses originating from different targets" and "two neighboring hypotheses originating from the same target." This invention creatively introduces "measurement source labels," assigning each Gaussian component a "pedigree" or "identity" of its information source, thereby establishing a two-dimensional joint criterion combining state correlation and source consistency. Based on this criterion, the "merging suppression" strategy can resolutely prevent the merging of two Gaussian components, even when they are physically highly overlapping, as long as they are determined to originate from different sensor measurements. This fundamentally solves the problem of nearby target loss caused by erroneous merging, significantly enhancing target resolution and tracking integrity in high-density scenes.

[0031] This invention discloses a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering, which introduces a "weight protection" mechanism based on newly generated labels, significantly enhancing the tracking robustness for emerging and weak targets. In complex tracking environments, the signal of a real target may be temporarily weakened or divided among multiple targets, resulting in low weights for its corresponding Gaussian components. These components are easily mistaken for clutter and incorrectly deleted by traditional fixed-threshold pruning strategies. The "weight protection" mechanism of this invention can intelligently identify low-weight components with "new" labels supported by new measurements at the current moment. The system grants these components a brief "exemption period," protecting them from immediate pruning and giving them the opportunity to prove their value through continuous observation in subsequent moments. This technical feature greatly improves the filter's detection sensitivity for emerging targets and the tracking stability for weak targets, effectively reducing the unexpected interruption rate of the track.

[0032] This invention presents a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering, achieving a leap from "static threshold" to "dynamic adaptive decision-making," thus improving the algorithm's intelligence and scene adaptability. The core of this invention is a dynamic, context-based decision-making process, rather than a fixed set of unchanging rules. It comprehensively judges and executes the most reasonable operation (merging, suppressing merging, pruning, or protecting) based on the real-time state, weight, and unique source label of each Gaussian component. This adaptive characteristic makes the entire tracking system more intelligent, better adapting to various complex and dynamically changing tracking scenarios, from sparse to dense, from long-distance to short-distance. Compared to traditional techniques that rely on manually setting and adjusting fixed thresholds, it exhibits stronger robustness and superior overall performance.

[0033] The multi-target tracking method based on Gaussian mixture probability hypothesis density filtering described in this invention effectively overcomes the key bottleneck of traditional GM-PHD filters in near-multi-target scenarios by introducing measurement source labels and establishing a novel adaptive decision logic, ultimately achieving a significant improvement in multi-target tracking accuracy, robustness, and continuity. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of a thermal insulation cover for protecting the temperature of an elevator traction machine from low temperatures, according to the present invention.

[0035] Figure 2 This is a schematic diagram of a scenario for near multi-target tracking provided by an embodiment of the present invention, wherein (a) is a near scenario of parallel targets and (b) is a near scenario of intersecting targets;

[0036] Figure 3 This is a diagram illustrating the effects of a traditional merger strategy.

[0037] Figure 4 This is a diagram illustrating the effect of the merging suppression strategy provided in the embodiments of the present invention;

[0038] Figure 5 This is a schematic diagram of the internal logic of the adaptive pruning and merging module provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0040] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0041] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 5 Detailed explanation is as follows:

[0042] Example 1:

[0043] A multi-target tracking method based on Gaussian mixture probability hypothesis density filtering includes the following steps:

[0044] S1. At each filtering time, acquire the measurement data corresponding to each sensor and assign the measurement identifier corresponding to the measurement data;

[0045] Furthermore, in each filtering cycle (time k), the sensor returns a set of N measurements;

[0046]

[0047] First, for each independent measurement in this group of measurements Assign a unique measurement identifier for that moment. In the subsequent GM-PHD filter update step, if a Gaussian component is measured... If an update is performed, the Gaussian component will inherit... As its own measurement source label, the Gaussian component retains its original label during the prediction step. In this way, each Gaussian component carries its identity information, i.e., which specific observation it is supported by.

[0048] S2. Based on the measurement data corresponding to each sensor obtained in step S1, update the Gaussian component of each sensor and set the measurement source label of the Gaussian component to the measurement identifier corresponding to the measurement data of that sensor.

[0049] Furthermore, in step S2, the i-th Gaussian component at time k includes the weight of the Gaussian component, the mean of the Gaussian component, the covariance of the Gaussian component, the measurement source label carried by the Gaussian component, and the timestamp when the measurement source label is assigned.

[0050] Furthermore, the Gaussian component, the basic operational unit in the traditional GM-PHD filter, is given an extended definition. A Gaussian component is defined as a unique mathematical structure with historical properties. :

[0051]

[0052] in, These are the weight, mean, and covariance of the i-th component at time k, respectively. This is the measurement source tag carried by the component, used to indicate its information source; It is the timestamp when the label was assigned, i.e., the "label age", which is used for subsequent weight protection decisions.

[0053] Specifically, regarding the problem of tracking multiple nearby objects, such as Figure 2 As shown, in scenarios such as autonomous vehicles and drone swarms, multiple targets (Target 1, Target 2) may be extremely close in physical space. When sensors (such as radar) observe them, the resulting measurement data (Measurement 1, Measurement 2) also show very close distances in the measurement space. Traditional methods rely solely on distance or state similarity for judgment, which can easily lead to the erroneous merging of hypotheses representing two different targets into one, resulting in the loss of tracking of one of the targets. The goal of this invention is to stably and accurately maintain the tracking of each individual target even in such close proximity scenarios.

[0054] S3. Establish an adaptive pruning and merging decision model;

[0055] Furthermore, the specific implementation method of step S3 includes the following steps:

[0056] S3.1. Traverse all Gaussian component pairs and calculate the state correlation of all Gaussian component pairs using the Bach distance; if the Bach distance is less than the merging distance threshold, proceed to the merging decision process; if the Bach distance is greater than the merging distance threshold, suppress merging and proceed to the next step.

[0057] The merging decision process is set to check the measurement source labels of Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are different, the merging process is suppressed and the next step is not performed. If the measurement source labels of the Gaussian component pairs are the same, the merging process is performed to generate merged Gaussian components.

[0058] In step S3.1, the i-th Gaussian component is calculated using the Bach distance. and the j-th Gaussian component The state correlation between them is expressed as:

[0059]

[0060] in, Let the Bach distance be the distance between the i-th Gaussian component and the j-th Gaussian component. and Let be the mean and covariance of the i-th and j-th Gaussian components, respectively. denoted as the average covariance of the i-th Gaussian component and the j-th Gaussian component; det is the value of the matrix determinant.

[0061] Furthermore, in step S3.1, the state correlation is combined with the measurement source label, and a merging decision function based on label dependence is used for judgment. Then, a merging operation is performed on two or more Gaussian components, or the merging operation on two or more Gaussian components is prohibited. The judgment formula is as follows:

[0062]

[0063] in, Let be the decision function for merging the i-th Gaussian component and the j-th Gaussian component. These are the measurement source labels for the i-th Gaussian component and the j-th Gaussian component, respectively. This is the merging threshold based on Bach distance.

[0064] S3.2. Perform pruning decisions on all merged Gaussian components and unmerged Gaussian components. If the weight of a Gaussian component is greater than the pruning weight threshold, no pruning is performed. If the weight of a Gaussian component is less than the pruning weight threshold, check the age and protection status of the measurement source labels of the Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are old time labels and are not within the protection period, pruning is performed. If the measurement source labels of the Gaussian component pairs are new time labels, no pruning is performed.

[0065] Furthermore, in step S3.2, a pruning decision is made using a dynamic pruning function P based on the labeled age. P outputs 1 to represent retention and 0 to represent pruning. The expression is:

[0066]

[0067] Where u is the unit step function, For component weights, The basic pruning threshold is k, where k is the time step. The labeled age for the component. This is the weight protection period.

[0068] S4. Input all Gaussian components into the adaptive pruning and merging decision model established in step S3, and perform merging or pruning to obtain the remaining Gaussian components.

[0069] Using the weights of the Gaussian components, the state correlations between the Gaussian components, and the measurement source labels carried by the Gaussian components as joint criteria, adaptive pruning and merging processing is performed on the set of Gaussian components in the PHD, so that the filter can distinguish and retain the independent hypothesis of neighboring targets. Specifically, its internal logic can be referred to Figure 5 As shown, it comprises two main sub-modules: merging and pruning. Its decisions no longer rely on fixed thresholds but are based on a dynamic, multi-dimensional joint criterion. In the merging implementation, the goal is to intelligently determine whether two Gaussian components with similar states should be merged, as shown in the comparison. Figure 3 and Figure 4 As shown.

[0070] Furthermore, the Gaussian components in the PHD are traversed; it is determined whether the weight of the Gaussian component is lower than a preset pruning weight threshold, and whether the measurement source label of the Gaussian component is a newly assigned label at the current filtering time; if the weight of the Gaussian component is lower than the pruning weight threshold, but its measurement source label is a newly assigned label at the current filtering time, then a weight protection operation is performed on the Gaussian component to prevent components representing newly emerging or reconfirmed real targets from being prematurely deleted.

[0071] Furthermore, the weight protection operation exempts the Gaussian component from pruning during the current filtering cycle; and applies a specific pruning weight threshold lower than the normal pruning weight threshold to the Gaussian component during the current filtering cycle.

[0072] S5. Set the state extraction weight threshold, compare the retained Gaussian components obtained in step S4, and output the mean of the Gaussian components with weights higher than the state extraction weight threshold as the final extracted target state.

[0073] Furthermore, after adaptive processing, the final posterior PHD intensity function is obtained. The posterior strength function, which can be clearly described by a label, highlights the unique ability of this invention to classify and manage information sources:

[0074]

[0075] In the formula, Represents all measured The updated version includes new tags. The sum of the intensities of the Gaussian components clearly demonstrates which different sources of "information flow" converge to form the final posterior distribution.

[0076] Furthermore, in the prediction step of the GM-PHD filter, the Gaussian component records and retains its measurement source label at all times during the guard period.

[0077] Furthermore, in this embodiment, by introducing measurement source labels, the management of Gaussian component assumptions is elevated from purely physical space analysis to a higher dimension that combines physical space with information sources. The entire decision-making process can intelligently distinguish between two fundamentally different situations: "appearing close" and "from the same source," enabling the filter to maintain extremely high tracking accuracy and robustness even in complex scenarios such as high density, high clutter, and near-target situations.

[0078] Furthermore, the apparatus of this embodiment includes: a PHD processing module, used to establish a PHD hypothesis management model that incorporates measurement source labels based on the calculation process of estimating the target state through a GM-PHD filter in multi-target tracking and pruning and merging the probability hypothesis density (PHD) in Gaussian mixture form output by the filter; wherein each Gaussian component in the PHD is assigned a measurement source label to characterize its update source; and an adaptive decision module, used to perform adaptive pruning and merging processing on the set of Gaussian components in the PHD using the weights of the Gaussian components, the state correlation between the Gaussian components, and the measurement source labels carried by the Gaussian components as joint criteria, wherein, based on the joint criteria, when two or more Gaussian components with state correlation higher than a preset threshold are identified as having different measurement source labels, a merging suppression operation is performed on these Gaussian components; when a Gaussian component with a weight lower than a preset threshold is identified as having a newly assigned label at the current time, a weight protection operation is performed on that Gaussian component.

[0079] This embodiment describes a multi-target tracking method based on Gaussian mixture probability hypothesis density filtering. The method includes establishing an adaptive pruning and merging decision model based on the process of estimating the target state using a GM-PHD filter and pruning and merging the Gaussian components output by the filter. Each Gaussian component in the PHD is assigned a measurement source label. Using the weights of the Gaussian components, the state correlation between Gaussian components, and the measurement source labels carried by the Gaussian components as joint criteria, adaptive pruning and merging processing is performed on the set of Gaussian components in the PHD, preserving the independent hypothesis of neighboring targets. During the adaptive pruning and merging process, two or more Gaussian components in the PHD with a state correlation higher than a first preset threshold are identified using the joint criteria. It is then determined whether the measurement source labels of the two or more Gaussian components are identical. If the measurement source labels are identical, a merging operation is performed on the two or more Gaussian components.

[0080] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0081] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A multi-target tracking method based on Gaussian mixture probability hypothesis density filtering, characterized in that, Includes the following steps: S1. At each filtering time, acquire the measurement data corresponding to each sensor and assign the measurement identifier corresponding to the measurement data; S2. Based on the measurement data corresponding to each sensor obtained in step S1, update the Gaussian component of each sensor and set the measurement source label of the Gaussian component to the measurement identifier corresponding to the measurement data of that sensor. S3. Establish an adaptive pruning and merging decision model; The specific implementation method of step S3 includes the following steps: S3.

1. Traverse all Gaussian component pairs and calculate the state correlation of all Gaussian component pairs using the Bach distance; if the Bach distance is less than the merging distance threshold, proceed to the merging decision process; if the Bach distance is greater than the merging distance threshold, suppress merging and proceed to the next step. The merging decision process is set to check the measurement source labels of Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are different, the merging process is suppressed and the next step is not performed. If the measurement source labels of the Gaussian component pairs are the same, the merging process is performed to generate merged Gaussian components. In step S3.1, the i-th Gaussian component is calculated using the Bach distance. and the j-th Gaussian component The state correlation between them is expressed as: in, Let the Bach distance be the distance between the i-th Gaussian component and the j-th Gaussian component. and Let be the mean and covariance of the i-th and j-th Gaussian components, respectively. is the average covariance of the i-th Gaussian component and the j-th Gaussian component; det is the value of the matrix determinant; In step S3.1, the state correlation is combined with the measurement source label, and a merging decision function based on label dependence is used for judgment. Then, a merging operation is performed on two or more Gaussian components, or the merging operation on two or more Gaussian components is prohibited. The judgment formula is as follows: in, Let be the decision function for merging the i-th Gaussian component and the j-th Gaussian component. These are the measurement source labels for the i-th Gaussian component and the j-th Gaussian component, respectively. The merging threshold is based on the Bach distance. S3.

2. Perform pruning decisions on all merged Gaussian components and unmerged Gaussian components. If the weight of a Gaussian component is greater than the pruning weight threshold, no pruning is performed. If the weight of a Gaussian component is less than the pruning weight threshold, check the age and protection status of the measurement source labels of the Gaussian component pairs. If the measurement source labels of the Gaussian component pairs are old time labels and are not within the protection period, pruning is performed. If the measurement source labels of the Gaussian component pairs are new time labels, no pruning is performed. S4. Input all Gaussian components into the adaptive pruning and merging decision model established in step S3, and perform merging or pruning to obtain the remaining Gaussian components. S5. Set the state extraction weight threshold, compare the retained Gaussian components obtained in step S4, and output the mean of the Gaussian components with weights higher than the state extraction weight threshold as the final extracted target state.

2. The multi-target tracking method based on Gaussian mixture probability hypothesis density filtering according to claim 1, characterized in that, In step S2, the i-th Gaussian component at time k includes the weight of the Gaussian component, the mean of the Gaussian component, the covariance of the Gaussian component, the measurement source label carried by the Gaussian component, and the timestamp when the measurement source label is assigned.

3. The multi-target tracking method based on Gaussian mixture probability hypothesis density filtering according to claim 2, characterized in that, In step S3.2, a pruning decision is made using a dynamic pruning function P based on the labeled age. P outputs 1 to represent retention and 0 to represent pruning. The expression is: Where u is the unit step function, For component weights, The basic pruning threshold is k, where k is the time step. The labeled age for the component. This is the weight protection period.

4. The multi-target tracking method based on Gaussian mixture probability hypothesis density filtering according to claim 3, characterized in that, Traverse the Gaussian components in the PHD; determine whether the weight of the Gaussian component is lower than a preset pruning weight threshold, and whether the measurement source label of the Gaussian component is a newly assigned label at the current filtering time; if the weight of the Gaussian component is lower than the pruning weight threshold, but its measurement source label is a newly assigned label at the current filtering time, then perform a weight protection operation on the Gaussian component to prevent components representing newly emerging or reconfirmed real targets from being prematurely deleted.

5. A multi-target tracking method based on Gaussian mixture probability hypothesis density filtering according to claim 4, characterized in that, The weight protection operation exempts the Gaussian component from pruning during the current filtering cycle; and applies a specific pruning weight threshold lower than the normal pruning weight threshold to the Gaussian component during the current filtering cycle.

6. The multi-target tracking method based on Gaussian mixture probability hypothesis density filtering according to claim 5, characterized in that, In the prediction step of the GM-PHD filter, the Gaussian component records and retains the measurement source label it has at all times during the guard period.

Images