A three-dimensional multi-target tracking method for complex category target accurate trajectory prediction

Abstract

The application discloses a three-dimensional multi-target tracking method for complex category target accurate trajectory prediction, relates to the technical field of trajectory prediction planning, and comprises a trajectory motion module, a multi-category data association module and a trajectory management module. A basic model and a two-round target motion model are constructed through a dynamic dissimilarity index classification, two-stage adaptive data association is realized in combination with FAIoU and Ro_GDIoU, and a confidence weighted penalty and a WNMS are adopted to optimize a trajectory management strategy. The application can significantly improve the tracking precision and robustness of the system in a complex dynamic environment, and is suitable for multi-category three-dimensional target tracking scenes such as automatic driving and intelligent traffic monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trajectory prediction planning, in particular to a three-dimensional multi-target tracking method for precise trajectory prediction of complex category targets. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle cluster cooperation, intelligent monitoring systems and automatic driving technologies, higher requirements are put forward for multi-target perception and accurate tracking in complex dynamic environments. As a core perception task in the above systems, three-dimensional multi-target tracking (3DMOT) aims to continuously identify and predict the three-dimensional trajectories of multiple heterogeneous targets in the scene based on lidar, image or multi-modal input data. However, the current mainstream methods still have the following technical problems in multi-category dynamic environments:

[0003] (1) Lack of category difference modeling mechanism: existing methods usually use a unified trajectory prediction model, such as constant acceleration model (CA) or constant velocity model (CV), which assumes that all target categories have similar motion behavior. This simplified modeling method ignores the essential differences in physical structure, turning radius and dynamic performance of targets, and cannot accurately describe the real motion rules of heterogeneous targets such as vehicles, pedestrians, bicycles and unmanned aerial vehicles.

[0004] (2) Insufficient tracking accuracy and stability: due to the above unified modeling strategy, which does not have the ability to distinguish between multiple categories of targets, it often leads to problems such as trajectory drift, large prediction deviation, etc. in multi-target dense or high dynamic scenes, thereby affecting the matching accuracy and trajectory continuity, resulting in high false positive rate (False Positives) and identity switching rate (IDSwitch).

[0005] (3) Lack of generalization ability for real complex environments: some methods attempt to introduce models such as CTRA or Bicycle to model individual categories, but such models rely on manually specified category structure rules, have limited scope of application, and are difficult to adapt to the problems of uncertain target categories, fast dynamic changes, complex multi-source inputs, etc. in actual scenes, which restricts their practicality and scalability in automatic driving and unmanned systems.

[0006] In summary, there is an urgent need for a three-dimensional multi-target tracking method with strong generalization ability and multi-target structure adaptability, which can establish a flexible and stable motion modeling mechanism for different target category differences, and combine high-precision data association and trajectory management strategies to meet the tracking needs of multi-source heterogeneous targets in real-world scenarios. SUMMARY

[0007] The application aims to provide a three-dimensional multi-target tracking method for complex category target accurate trajectory prediction, mainly solving the technical problems of insufficient category difference modeling, lacking of tracking accuracy and stability, and limited adaptability to real complex environment of the existing three-dimensional multi-target tracking method in complex dynamic environment.

[0008] To achieve the above-mentioned purpose, the technical scheme adopted by the application is as follows:

[0009] A three-dimensional multi-target tracking method for complex category target accurate trajectory prediction, which is realized based on a PADMOT framework, the PADMOT framework including a trajectory motion modeling module, a multi-category data association module, and a trajectory management module; the specific implementation steps are as follows:

[0010] S1, accurately modeling the diversified motion behavior of the target object by using the trajectory motion modeling module;

[0011] S2, constructing a FAIoU similarity measurement index by using the multi-category data association module, and realizing the association of the trajectory and the detection result of the target object by using a two-stage matching strategy;

[0012] S3, guaranteeing the robustness and real-time performance of the trajectory prediction of the target object by using the trajectory management module based on a dynamic feature trajectory management strategy.

[0013] Further, the specific process of the step S1 is as follows:

[0014] S11, classifying the target object based on a dynamic dissimilarity index DDI, and dividing the target object into a basic structure and a two-wheel structure;

[0015] S12, for the basic structure, adopting a standard motion model, combining constant acceleration and steering rate to construct three state vectors and respectively modeling; for the two-wheel structure, adopting a double-wheel motion model, introducing a steering angle to establish a heading deviation relationship, and constructing a state vector coupled with the heading angle;

[0016] S13, decoupling the linear and nonlinear variables in the state vector, and respectively applying a Kalman filter for independent prediction update.

[0017] Further, the specific process of the step S2 is as follows:

[0018] S21, constructing a FAIoU similarity measurement index including a center distance, a size difference, and a heading angle penalty term;

[0019] S22, in the first stage, calculating the matching similarity of the last frame trajectory and the current detection result based on the FAIoU;

[0020] S23, in the second stage, Ro_GDIoU metric is introduced in the bird's eye view perspective plane, and secondary matching is performed on the unmatched targets in the first stage;

[0021] S24, the correlation robustness is improved and false positive matching and ID switching are reduced through the two-stage matching strategy.

[0022] Further, the specific process of step S3 is as follows:

[0023] S31, a judgment rule based on speed and acceleration change is introduced in trajectory initialization, which is used to filter effective candidate targets and suppress false targets;

[0024] S32, in trajectory updating, a barycenter geometric adjustment strategy is adopted to calculate the dynamic geometric center of two rounds of targets to improve updating accuracy;

[0025] S33, the max-age threshold of the trajectory life cycle is dynamically adjusted according to the motion state of the target object to adapt to the tracking needs of targets with different speeds;

[0026] S34, a confidence weighted penalty mechanism is introduced, which strengthens the punishment of trajectories below the minimum confidence threshold, and guarantees the stability of trajectories that meet the set confidence threshold;

[0027] S35, in the trajectory output stage, a weighted non-maximum suppression mechanism is adopted to suppress overlapping trajectories according to the confidence, and the final effective result is output.

[0028] Further, in the step S11, the dissimilarity index DDI includes the average angle between target speed and heading and normalized speed, representing the motion consistency and maneuverability of the target, wherein:

[0029] (1)

[0030] wherein, represents the average angular deviation, which is obtained by measuring the average angle between the speed vector and the rotation direction; represents the average speed, which is in m / s and serves as a proxy for target maneuverability and dynamic ability; is the maximum average speed observed in all categories, which is used for normalization.

[0031] Further, in the step S12, the expression of the standard motion model is:

[0032] (2)

[0033] wherein, , , respectively represent the position state vector, the size state vector and the angle state vector of the target object in the standard motion model; represent the geometric center coordinates of the target object, v x 、v y respectively represent the velocity components of the target object in the x-axis direction and the y-axis direction, a x 、a y respectively represent the acceleration components of the target object in the x-axis direction and the y-axis direction, represent the width, height and length of the target object; v w 、v l 、 v h respectively represent the velocity components of the target object in the width, height and length directions; and the variables correspond to the heading angle and the turning rate, respectively; represent the heading angle obtained from the detection, represent the heading angle calculated according to the velocity, ; ω d represent the turning rate obtained from the detection, ω v represent the turning rate calculated according to the velocity.

[0034] Further, the expression of the double-wheel motion model is:

[0035] (3)

[0036] wherein, 、 、 respectively represent the position state vector, the size state vector and the angle state vector of the target object in the double-wheel motion model, represent the center of mass coordinates of the target object, v x’ 、v y’ respectively represent the velocity components of the target object in the x-axis direction and the y-axis direction in the double-wheel motion model, a x’ 、a y’ respectively represent the acceleration components of the target object in the x-axis direction and the y-axis direction in the double-wheel motion model; represent the turning angle of the target object.

[0037] Further, in the step S22, the calculation process of the matching similarity is as follows:

[0038] S221, let denote the track in the previous frame, denote the result of detection in the current frame, each frame uses FAIoU metric to fuse multiple three-dimensional geometry related items;

[0039] S222, introduce multiple nonlinear normalization processes to capture subtle differences between highly similar objects and accurately identify the correct object;

[0040] S223, after initializing the data, use FAIoU to calculate and the similarity matrix between them;

[0041] S224, after obtaining the result, use the Hungarian matching algorithm to associate with ;

[0042] S225, determine the association threshold specific to the iteration test to implement the constraint of class matching;

[0043] S226, derive the matched detection , matched track , unmatched detection and unmatched track as the input of the second stage association.

[0044] Further, in the step S34, the confidence weighted penalty mechanism applies different magnitudes of exponential penalty to tracks with different confidences using a confidence weighted penalty function, enhancing the track retention selectivity.

[0045] Further, in the step S35, the weighted non-maximum suppression mechanism filters and suppresses overlapping targets in the output result according to track confidence, improving the accuracy and stability of the output.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] (1) The PADMOT framework proposed in the present application, which comprises a track motion module, a multi-class data association module and a track management module, can be widely applied in automatic driving perception systems, intelligent traffic monitoring, multi-robot collaborative perception, crowd behavior recognition and other application scenarios with high requirements for multi-class three-dimensional target tracking.

[0048] (2) The application divides the target into basic structure and two-wheel structure through dynamic dissimilarity index (DDI), respectively adopts standard motion model (combined with constant acceleration and steering rate) and double-wheel motion model (introducing steering angle and heading angle coupling), solves the problem of insufficient description of heterogeneous target motion law in the prior art, and improves the accuracy of multi-class target motion modeling.

[0049] (3) The application adopts two-stage data association strategy (FAIoU similarity measurement + bird's eye view Ro_GDIoU secondary matching), combines confidence weighted penalty mechanism and weighted non-maximum suppression (WNMS), effectively reduces false positive matching and ID switching rate, enhances trajectory continuity, and solves the tracking drift problem of the traditional method in the multi-target dense or high dynamic scene.

[0050] (4) The application optimizes the trajectory update by dynamically adjusting the trajectory life cycle threshold (max-age) and introducing the center of gravity adjustment (CGA) strategy, and improves the robustness to complex scenes such as rapid movement and temporary occlusion, and overcomes the defects that the existing method has limited generalization ability to the real environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a trajectory motion module framework in the method of the application.

[0052] Figure 2 It is a multi-class data association module framework in the method of the application.

[0053] Figure 3 It is a trajectory management module framework in the method of the application.

[0054] Figure 4 It is a trajectory prediction result comparison diagram of two-wheel objects under different motion models in the method of the application.

[0055] Figure 5 It is a 2D visualization diagram of the two-wheel motion model and other model prediction performance in the method of the application.

[0056] Figure 6 It is a PADMOT trajectory overhead view (basic target) under scene 1040 in the embodiment of the application.

[0057] Figure 7 It is a MCTrack trajectory overhead view (basic target) under scene 1040 in the embodiment of the application.

[0058] Figure 8 It is a PADMOT trajectory overhead view (two-wheel target) under scene 1032 in the embodiment of the application.

[0059] Figure 9This is a top-down view of the MCTrack trajectory (two-wheel target) in scenario 1032 of this embodiment of the invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0061] This invention discloses a 3D multi-target tracking method for accurate trajectory prediction of complex target categories. This method is implemented based on the PADMOT framework, such as... Figures 1-3 As shown, PADMOT consists of a trajectory motion module, a multi-category data association module, and a trajectory management module. This invention can be implemented based on mainstream perception system architectures, combining LiDAR point clouds and camera image inputs, and using the aforementioned modules for parallel or distributed tracking processing. Specific parameters can be set and fine-tuned according to different target categories and scene requirements. The specific implementation can be constructed by referring to general filter libraries and data association modules in this field.

[0062] The method of the present invention includes the following steps:

[0063] Step 1: Use the trajectory motion modeling module to accurately model the diverse motion behaviors of the target object.

[0064] Step 11: To characterize the fundamental differences in dynamic behavior among object categories, this embodiment introduces a quantitative metric based on the Dynamic Dissimilarity Index (DDI). This index simultaneously captures the degree of dynamic movement of each category, decoupling and making it operable. The metric defines the Dissimilarity Index (DDI):

[0065] (1)

[0066] in, The average angular deviation is obtained by measuring the average angle between the velocity vector and the direction of rotation. It represents the average speed, measured in m / s, and serves as a proxy for the target's maneuverability and dynamic capabilities. It is the maximum average velocity observed across all categories, used for normalization.

[0067] Step 12: For the basic structure, a standard motion model is adopted, and three state vectors are constructed by combining constant acceleration and steering rate, and modeled separately. For the two-wheel structure, a two-wheel motion model is adopted, introducing the steering angle to establish the heading deviation relationship, and constructing a state vector coupled with the heading angle. For example... Figure 4Shown: Comparison of trajectory prediction results for two-wheel objects using different motion models. The CA model exhibits a clear bias due to its inability to account for turning behavior, assuming constant acceleration along the current direction. The CTRA model and the bicycle model show improved performance by capturing the basic turning dynamics. However, they still produce a noticeable bias as they lack explicit modeling of the decoupling relationship between the standard direction and velocity. To address this issue, the invention separates the state vector and applies separate Kalman filters to the linear and nonlinear components, thereby improving stability and efficiency.

[0068] The standard motion model is suitable for structurally stable base targets, including standard objects such as cars, pedestrians, buses, trailers, and trucks, with the specific formula as follows:

[0069] (2)

[0070] where, , , denote the position state vector, size state vector, and angle state vector of the target object in the standard motion model, respectively; denotes the geometric center coordinates of the target object, v x 、v y denote the velocity components of the target object in the x-axis and y-axis directions, respectively, a x 、a y denote the acceleration components of the target object in the x-axis and y-axis directions, respectively, denote the width, height, and length of the target object; v w 、v l , v h denote the velocity components of the target object in the width, height, and length directions, respectively; and variables correspond to the heading angle and the turning rate, respectively; denotes the heading angle obtained from detection, denotes the heading angle calculated from velocity, ; ω d denotes the turning rate obtained from detection, ω v denotes the turning rate calculated from velocity.

[0071] Step 13, In practical situations, the coordinates of the target object can change, while its size usually remains unchanged. However, since the sum of the changes along the axes is small and has a negligible impact on the prediction, this embodiment optimizes the corresponding indicators in this direction through the filter. The double-wheel motion model is suitable for two-wheeled targets with speed and heading decoupling, including bicycles and motorcycles, using constant steering angle and acceleration prediction, while using constant speed prediction and the present invention uses three state vectors to represent the trajectory state of the object, as follows:

[0072] (3)

[0073] wherein, , , denote the position state vector, size state vector and angle state vector of the target object in the double-wheel motion model, respectively, denotes the centroid coordinates of the target object, v x’ 、v y’ denote the velocity components of the target object in the x-axis direction and y-axis direction in the double-wheel motion model, respectively, a x’ 、a y’ denote the acceleration components of the target object in the x-axis direction and y-axis direction in the double-wheel motion model, respectively; denotes the steering angle of the target object.

[0074] Step 14, The double-wheel motion model aims to overcome these limitations. For two-wheeled objects such as bicycles and motorcycles, it explicitly incorporates the steering angle and its relationship with the heading angle direction to better capture their unhooked motion patterns. In addition, this embodiment decomposes the high-dimensional state vector into three interpretable components and applies separate Kalman filters to each filter, thereby enhancing the prediction stability and accuracy. As Figure 5 shown, this method proposed by the present invention can significantly improve the trajectory estimation of various object types and provide stronger protection for real-world autonomous driving scenarios.

[0075] Step 2, As Figures 1-3 shown, the present invention designs a multi-class data association module that applies two different similarity indicators at different stages. FAIoU is proposed for the first association, which contains multiple 3D information penalty terms to achieve adaptive similarity measurement for different object classes, thereby improving the accuracy and robustness of data association.

[0076] Step 21, FAIoU similarity measurement indicators are constructed, including center distance, size difference and heading angle penalty terms.

[0077] Step 22, in the first stage, the matching similarity between the previous frame trajectory and the current detection result is calculated based on FAIoU, aiming to maximize the alignment between the trajectory and the detection result.

[0078] The embodiment proposes FAIoU, which integrates a penalty term to adjust the spatial positioning and geometric correlation, especially the center distance and size change. This provides a simple and effective method to adapt to processing multi-class objects. The embodiment not only relies on 3D information, but also integrates multiple object similarities into the penalty term. FAIoU is represented as follows:

[0079] (4)

[0080] Where G represents GIoU, and D represents DIoU; represents 3DIoU, forms a penalty term, which integrates the position and geometric correlation into the similarity measurement. represents the constraint weight of the penalty, aiming to avoid false positive association.

[0081] To capture subtle differences between highly similar objects and accurately identify the correct object, multiple nonlinear normalization processes are introduced:

[0082] (5)

[0083] (6)

[0084] (7)

[0085] (8)

[0086] (9)

[0087] (10)

[0088] Where, and represent the 3D bounding box of the detection and the trajectory, respectively, C represents the closed convex hull , c represents the diagonal length of the smallest closed box; represents the bounding box parameters of the detection and the trajectory, including width, length, height and heading angle; represents the absolute difference operation.

[0089] After initializing the data, FAIoU is used to calculate the previous trajectory) and Similarity matrix between (current detection). To further improve the continuity and robustness of the trajectory, the first association is performed using bidirectional prediction fusion. Based on the proposed DDI-based motion model, the Kalman filter is used to predict the trajectory while retrospectively inferring the detection results. The trajectory of the t-th frame is denoted as xt, and the trajectory prediction at this time is defined as:

[0090] (11)

[0091] and represent the state transition functions of the standard and two-wheeled objects, respectively, which can predict the state change between two observations. denotes the predicted trajectory in the previous frame. In this process, both the standard object and the two-wheeled object conform to constant acceleration. Essentially, the geometric center coordinates and the center of mass coordinates do not need to be differentiated in practice. The position conversion process of the target is represented as follows:

[0092] (12)

[0093] (13)

[0094] where and are the trajectory prediction coordinates of the t-1 frame, is the laser radar frequency. According to their respective motion models, different position conversion processes are established for the standard and two-wheeled objects. At this time, the time interval is defined as the distance between the integral variables and the lower limit of integration in the integration process . The equation is as follows:

[0095] (14)

[0096] (15)

[0097] (16)

[0098] (17)

[0099] The detection result at the conversion framework based on is denoted as , and the retrospective inference result is given by the following way:

[0100] (18)

[0101] where This represents the state transition function. Both the base objective and the two rounds of objectives in the process use a constant velocity model. After completing the above calculations, the detection results (actual detection results) will be calculated. and traceability testing ) and trajectory (known trajectory) and predicted trajectory The cost function between () is given by the following formula:

[0102] (19)

[0103] in, This represents the FAIoU similarity metric. Bidirectional prediction fusion leverages the advantages of both forward and backward predictions to achieve more accurate, smooth, and coherent trajectory estimation. The results are then processed using the Hungarian matching algorithm. and The association is then determined through iterative testing to identify category-specific association thresholds. This is used to enforce category matching constraints. Finally, a matching detection is obtained. Matching trajectory Unmatched detection and unmatched trajectories , as input for the second stage of association.

[0104] Step 23: This embodiment does not rely on threshold adjustments to achieve a minor improvement in accuracy. Instead, it prioritizes minimizing false positive matches and identity switching during the process. To achieve this, this embodiment proposes an association strategy that performs secondary matching in the BEV (Bird's-eye View Plane). This association process is designed for practical applications, relaxing the matching restrictions to better adapt to real-world challenges.

[0105] Step 24: In this embodiment, secondary association of previously unmatched detections and trajectories is performed by calculating RO_GDIOU. The final association result of the current framework is obtained using the Hungarian matching algorithm and a secondary association threshold.

[0106] Step 3, as follows Figure 1 As shown, this embodiment proposes a trajectory management module, which uses a trajectory management strategy based on dynamic features to ensure the robustness and real-time performance of the system in complex dynamic environments.

[0107] Step 31, Trajectory Initialization: Most previous algorithms initialize new detected trajectories as candidate trajectories based on a counting mechanism. However, relying solely on counting trajectory initialization can lead to false objects, especially when objects move quickly or experience partial occlusion. In the initialization process of the present embodiment, the present embodiment considers the motion state (e.g., velocity and acceleration). If the motion characteristics (e.g., velocity change) of an object exceed a predetermined threshold for multiple consecutive frames, it is more likely to be a valid object, and candidate trajectory initialization is prioritized. This approach effectively reduces the impact of noise and false objects.

[0108] Step 32, Trajectory Update: Traditional methods use the detected Dt and a standard extended Kalman filter (EKF) to update the trajectory state and covariance matrix. The EKF update process is performed through a state measurement transformation function. Previous methods assume a fixed geometric center of the object, usually based on its bounding rectangle or enclosing box. However, the actual geometric center (center of mass) of a two-wheeled object often changes due to changes in shape and posture. In the state measurement transformation function, the present embodiment introduces a center of mass-gravity adjustment (CGA) to determine the geometric center of the object. This means that for two-wheeled objects, the present embodiment no longer relies solely on the geometric center of the bounding box. Instead, the present embodiment updates the trajectory estimate according to the actual shape and center of mass of the object, thereby significantly improving tracking accuracy, especially in dynamic scenarios such as turning, acceleration, and deceleration.

[0109] Step 33, Trajectory Termination: Uniformly applying a fixed maximum age on all objects can lead to false removal of slow-moving objects or temporarily occluded objects. The present embodiment adaptively adjusts the maximum threshold according to the motion characteristics of the object. For fast-moving objects, the maximum age threshold is increased to prevent premature object loss, while for slow-moving or intermittently lost objects, it is reduced to quickly delete invalid trajectories.

[0110] Step 34, Trajectory Confidence Penalty: Current state-of-the-art methods use a penalty factor and an exponential function to penalize the confidence of retained trajectories. A simple exponential penalty mechanism applies uniform penalties to all weak trajectories, ignoring the penalty for high confidence. The present embodiment proposes a weighted penalty function that imposes stronger penalties on low-confidence trajectories while imposing weaker penalties on high-confidence trajectories. This weighted strategy ensures better retention of reliable trajectories and reduces the impact of false trajectories on overall system performance.

[0111] Step 35, result output: output contains all active tracks, updated track state, newly initialized tracks and part of unlabeled tracks. Weighted non-maximum suppression (WNMS) is used to reduce false positives. In the final output, the embodiment not only outputs the track state, but also annotates the confidence level and timestamp of each track. This helps subsequent track combination and analysis. The penalized track is accompanied by a confidence change case for further analysis and adjustment. The embodiment introduces a WNMS mechanism to list the priority of overlapping tracks as their confidence scores. This ensures that high point tracks are retained, and in the case of overlap, low confidence tracks are suppressed, effectively reducing false positives.

[0112] A large number of quantitative and qualitative experiments on the benchmark dataset show that the PADMOT model achieves the most advanced performance and better tracking results, as shown in the following table (Motorcycle ID, Bicycle ID, Pedestrian ID, Trailer ID, Truck ID, Bus ID in the figure represent the number of corresponding experimental objects during the experiment), respectively under scene 1040 (experiment-defined), scene 1032 (experiment-defined), the track effect of the basic target and the two-wheel target is compared by using the existing MCTrack and the PADMOT framework of the present application, which proves the superiority and necessity of the present application. Figures 6-9

[0113] The above embodiment is only one of the preferred embodiments of the present application and should not be used to limit the protection scope of the present application, but any modification or polishing without substantial meaning made within the main design idea and spirit of the present application, which still solves the technical problems consistent with the present application, should be included in the protection scope of the present application.​

Claims

1. A three-dimensional multi-target tracking method for complex category target precise trajectory prediction, characterized in that, The method is implemented based on a PADMOT framework, which comprises a trajectory motion modeling module, a multi-category data association module, and a trajectory management module; the specific implementation steps are as follows: S1, accurately modeling the diversified motion behavior of the target object by using the trajectory motion modeling module; the specific steps are as follows: S11, classifying the target object based on a dynamic dissimilarity index DDI, and dividing the target object into a basic structure and a two-wheel structure; S12, for the basic structure, adopting a standard motion model, combining constant acceleration and steering rate to construct three state vectors and modeling them respectively; for the two-wheel structure, adopting a double-wheel motion model, introducing a steering angle to establish a heading deviation relationship, and constructing a state vector coupled with the heading angle; S13, decoupling the linear and nonlinear variables in the state vector, and respectively applying Kalman filters for independent prediction update; S2, constructing a FAIoU similarity measurement index by using the multi-category data association module, and realizing the association of the trajectory and the detection result of the target object by a two-stage matching strategy; the specific steps are as follows: S21, constructing a FAIoU similarity measurement index including a center distance, a size difference, and a heading angle penalty term; S22, in the first stage, calculating the matching similarity of the last frame trajectory and the current detection result based on FAIoU; S23, in the second stage, introducing a Ro_GDIoU measurement in the bird's eye view plane to perform secondary matching on the target that is not matched in the first stage; S24, improving the association robustness and reducing false positive matching and ID switching by the two-stage matching strategy; S3, using the trajectory management module to guarantee the robustness and real-time performance of the trajectory prediction of the target object based on the trajectory management strategy of dynamic characteristics. 2.The three-dimensional multi-target tracking method for complex category target precise trajectory prediction of claim 1, wherein, The specific steps of the step S3 are as follows: S31, introducing a judgment rule based on speed and acceleration change in trajectory initialization, which is used for screening effective candidate targets and suppressing false targets; S32, adopting a barycenter geometric adjustment strategy in trajectory update to calculate the dynamic geometric center of the two-wheel target to improve the update accuracy; S33, dynamically adjusting the max-age threshold of the trajectory life cycle according to the motion state of the target object to adapt to the tracking needs of targets with different speeds; S34, introducing a confidence weighted penalty mechanism to strengthen the punishment of trajectories below the minimum confidence threshold, and guaranteeing the stability of trajectories reaching the set confidence threshold; S35, in the trajectory output stage, adopting a weighted non-maximum suppression mechanism to suppress overlapping trajectories according to the confidence, and outputting the final effective result. 3.The three-dimensional multi-target tracking method for complex category target precise trajectory prediction of claim 2, wherein, In the step S11, the dissimilarity index DDI includes the average angle between the target speed and the heading and the normalized speed, which represents the motion consistency and maneuvering ability of the target, wherein: (1) where, denotes the average angular deviation, obtained from the average angle between the measured velocity vector and the rotational direction; denotes the average velocity in m / s and serves as a proxy for the target maneuverability and dynamic capability; is the maximum average velocity observed in all classes and used for normalization.

4. The three-dimensional multi-target tracking method for complex category target precise trajectory prediction according to claim 3, characterized in that, In the step S12, the expression of the standard motion model is: (2) wherein , , respectively denote the position state vector, the size state vector and the angle state vector of the target object in the standard motion model; denotes the geometric center coordinate of the target object, v x 、v y respectively denote the velocity components of the target object in the x- and y- axis direction, a x 、a y respectively denote the acceleration components of the target object in the x- and y- axis direction, denotes the width, height and length of the target object; v w 、v l , v h respectively denote the velocity components of the target object in the width, height and length direction; and the variables correspond to the heading angle and the turn rate, respectively; denotes the heading angle obtained from the detection, denotes the heading angle calculated from the velocity, ; denotes the turn rate obtained from the detection, denotes the turn rate calculated from the velocity.

5. The three-dimensional multi-target tracking method for complex category target precise trajectory prediction according to claim 4, characterized in that, The expression of the double-wheel motion model is: (3) wherein, , , respectively represent a position state vector, a size state vector and an angle state vector of the target object in the double-wheel motion model, represents a mass center coordinate of the target object, v x’ , v y’ respectively represent a velocity component of the target object in the x-axis direction and the y-axis direction in the double-wheel motion model, a x’ , a y’ respectively represent an acceleration component of the target object in the x-axis direction and the y-axis direction in the double-wheel motion model; represents a steering angle of the target object.

6. The three-dimensional multi-target tracking method for complex category target precise trajectory prediction according to claim 5, characterized in that, In the step S22, the calculation process of the matching similarity is as follows: S221, let representing a trajectory in the previous frame, representing a result of detection in the current frame, each frame using a FAIoU metric to fuse multiple three-dimensional geometry related terms; S222, introducing multiple nonlinear normalization processes to capture subtle differences between highly similar objects and accurately identify the correct objects; S223, after initializing the data, compute the similarity matrix between the FAlOU using and the similarity matrix between the FAlOU using S224, after obtaining the result, the Hungarian matching algorithm is used to associate with the relevant. S225, determining a class-specific association threshold by iterative testing to enforce constraints on class matching; S226, deriving matched detections , matched tracks , unmatched detections and unmatched tracks as input to the second stage association.

7. The three-dimensional multi-target tracking method for complex category target precise trajectory prediction according to claim 6, characterized in that, In the step S34, the confidence-weighted penalty mechanism applies exponential penalty of different magnitudes to trajectories of different confidences using a confidence-weighted penalty function, enhancing the trajectory retention selectivity.

8. The three-dimensional multi-target tracking method for complex category target precise trajectory prediction according to claim 7, characterized in that, In the step S35, the weighted non-maximum suppression mechanism filters and suppresses overlapping targets in the output result according to the trajectory confidence, improving the accuracy and stability of the output.

Images