A 3D vision multi-target tracking method and system based on vehicle motion constraint

Abstract

The application discloses a 3D vision multi-target tracking method and system based on vehicle motion constraint, and belongs to the field of behavior recognition, and comprises the following steps: a six-degree-of-freedom motion model is used to predict a target object, and a target motion prediction trajectory is obtained; a data association of the target motion prediction trajectory and the target object is constructed, a Hungarian algorithm is used to solve the data association, and a target life cycle management auxiliary data association is used; the target motion prediction trajectory, the target object, the vehicle motion constraint and prior information are fused through a sliding window to construct a factor graph optimization model based on the sliding window; the motion state of the target object is optimized by using the factor graph optimization model based on the sliding window, and the optimized result is input into the six-degree-of-freedom motion model, and finally the motion trajectory of the target object is output. The application enhances the pose estimation accuracy and continuity of the target in a complex environment, and realizes the robust 3D multi-target tracking performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent unmanned systems, and in particular relates to a 3D visual multi-target tracking method and system based on vehicle motion constraints. Background Art

[0002] Multi-target tracking is a crucial component of the perception module of intelligent unmanned systems and is widely used in fields such as autonomous driving, intelligent surveillance, and intelligent robotics. Robust and reliable multi-target tracking provides autonomous unmanned systems with perceptual information about each target in a navigation scenario (primarily including identity, location, direction, and velocity), thereby assisting intelligent decision-making and planning and enabling safe autonomous navigation. Compared to 2D multi-target tracking methods, 3D multi-target tracking offers more comprehensive spatial perception capabilities. Furthermore, vision-based 3D multi-target tracking solutions have attracted widespread attention due to their low cost, rich texture information, and ease of installation.

[0003] Existing 3D visual multi-target tracking methods typically use Kalman filters or their variants to estimate the motion state of the tracked targets, and their state update mechanisms often rely solely on the target's detection bounding box information. However, in complex scenes, detection performance cannot be guaranteed for mutually occluded or distant targets, severely reducing the accuracy of target state estimation and trajectory prediction in existing methods.

[0004] Therefore, it is necessary to design a 3D visual multi-target tracking method and system based on vehicle motion constraints to address the above problems. Summary of the Invention

[0005] The purpose of the present invention is to address the problem of low target state estimation and trajectory prediction accuracy in existing methods, and to provide a 3D visual multi-target tracking method based on vehicle motion constraints, establish a six-degree-of-freedom motion model of the target, accurately predict the target trajectory to guide target detection bounding box matching, and reduce target confusion and false association; it is proposed to extend the vehicle motion constraints to 3D visual multi-target tracking, establish zero-speed constraints and non-holonomic constraints on the target trajectory state, enhance the pose estimation accuracy and continuity in complex environments, and achieve robust 3D multi-target tracking performance.

[0006] According to one aspect of this specification, a 3D visual multi-target tracking method based on vehicle motion constraints is provided, comprising:

[0007] S1. Use the six-degree-of-freedom motion model to predict the target object and obtain the target motion prediction trajectory;

[0008] S2. Construct data association between target motion prediction trajectory and target object, use Hungarian algorithm to solve data association, and use target lifecycle management to assist data association;

[0009] S3, fusing the target motion prediction trajectory, target object, vehicle motion constraints, and prior information through a sliding window to construct a sliding window-based factor graph optimization model;

[0010] S4. Use a sliding window-based factor graph optimization model to optimize the motion state of the target object, input the optimized result into the six-degree-of-freedom motion model, and finally output the motion trajectory of the target object.

[0011] Furthermore, the S2 includes:

[0012] Represent the target object as a target 3D detection bounding box;

[0013] Based on the target 3D detection bounding box and the target prediction trajectory, the association loss matrix is ​​constructed and solved using the Hungarian algorithm to obtain matching pairs. The expression of the association loss matrix is:

[0014]

[0015] in, represents the association loss between the i-th detected target and the j-th tracked target, Indicates whether the i-th detected target matches the j-th tracked target.

[0016] Furthermore, the vehicle motion constraints in S3 include zero-speed constraints and non-holonomic constraints:

[0017] The zero-speed constraint refers to constraining the parameters to be estimated by utilizing the physical characteristics of the vehicle when it is stationary, and constructing the zero-speed constraint residual in combination with the target posture difference;

[0018] The nonholonomic constraint refers to the assumption that the vehicle does not skid, drift, or bounce during driving, and the lateral and vertical velocities of the vehicle are zero. The nonholonomic constraint residual is constructed based on the orientation of each axis of the target coordinate system.

[0019] Furthermore, the expression for optimizing the motion state of the target object in S4 is:

[0020]

[0021] in, is the Cauchy robust kernel function, is the prior information obtained by marginalization, is the target motion prediction trajectory residual, is the target 3D detection bounding box residual, is the zero-velocity constraint residual, is the non-holonomic constraint residual, O is the set of all targets, M is the set of target motion model measurements, is the set of all stationary targets, is the set of stationary target motion model measurements, and B is the set of target 3D detection bounding box measurements.

[0022] Furthermore, after optimizing the motion state of the target object using the sliding window-based factor graph optimization model in S4, the step further includes: marginalizing the observation frames removed after optimization as prior information.

[0023] Furthermore, after obtaining the matching pair, the method further includes:

[0024] Based on the historical motion trajectory of failed matching, a six-degree-of-freedom motion model is used to predict the state and match again;

[0025] Based on the successfully matched matching pairs, they are input into the sliding window-based factor graph optimization model for motion state optimization;

[0026] Based on the detected target that fails to match, the detected target is marked as a new target.

[0027] Furthermore, the six-degree-of-freedom motion model in S1 includes six-degree-of-freedom constant acceleration and angular velocity, which is used to predict the motion trajectory of the target object.

[0028] According to one aspect of this specification, a 3D visual multi-target tracking system based on vehicle motion constraints is provided, comprising:

[0029] The trajectory prediction module is used to predict the target object using a six-degree-of-freedom motion model to obtain the target motion prediction trajectory;

[0030] The data association module is used to construct the data association between the target motion prediction trajectory and the target object, using the Hungarian algorithm to solve the data association and using the target lifecycle management to assist the data association;

[0031] The optimization model construction module is used to fuse the target motion prediction trajectory, target object, vehicle motion constraints and prior information through a sliding window to build a sliding window-based factor graph optimization model;

[0032] The target tracking module is used to optimize the motion state of the target object using a factor graph optimization model based on a sliding window, input the optimized results into a six-degree-of-freedom motion model, and finally output the motion trajectory of the target object.

[0033] According to one aspect of this specification, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the 3D visual multi-target tracking method based on vehicle motion constraints when executing the computer program.

[0034] According to one aspect of the present specification, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the 3D visual multi-target tracking method based on vehicle motion constraints are implemented.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The embodiment of the present invention proposes to extend vehicle motion constraints to 3D visual multi-target tracking, establish zero-speed constraints and non-holonomic constraints on the target trajectory state, enhance the accuracy and continuity of pose estimation in complex environments, and achieve robust 3D multi-target tracking performance.

[0037] 2. The embodiment of the present invention reduces target confusion and erroneous association by establishing a six-degree-of-freedom motion model for data association. By constructing a sliding window-based factor graph optimization model, it achieves joint and accurate estimation of the motion states of multiple targets, thereby improving the trajectory prediction and tracking accuracy of multiple targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 This is an algorithm framework diagram of an embodiment of the present invention;

[0040] Figure 2 A factor graph constructed for an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, an embodiment of the present invention provides a 3D visual multi-target tracking method based on vehicle motion constraints, including: S1, performing target detection based on GPS / IMU observation data and camera observation data to obtain a target 3D detection bounding box;

[0043] S2, a six-degree-of-freedom motion model is established to predict the state of the target, and a target motion prediction trajectory is obtained; the state of the target is projected to the camera coordinate system and compared with the target 3D detection bounding box, and a target 3D detection bounding box residual is obtained; S3, the Hungarian algorithm is used to match the target 3D detection bounding box and the target motion prediction trajectory, and when the matching is successful, a target motion prediction trajectory residual is obtained, which is combined with the target 3D detection bounding box residual and the vehicle motion constraint to construct a sliding window-based factor graph optimization model; S4, the target state is optimized by using the sliding window-based factor graph optimization model, and the joint accurate estimation of the multi-target state is realized, and the target trajectory prediction and tracking accuracy is improved.

[0044] Specifically, the target objects in step 1 are all represented by 3D detection bounding boxes, and the expression is as follows:

[0045] (1)

[0046] wherein, is the detection box corresponding to the i-th target in the camera coordinate system c, , are the three-dimensional position and attitude of the i-th target in the camera coordinate system c, is the length, width and height of the detection box corresponding to the i-th target. In order to prevent singularity problem, the attitude of the target in the method is represented in the form of quaternion.

[0047] In order to realize accurate and reliable target data association, the target detection box is converted to the world coordinate system w. Similar to the representation of the target in the camera coordinate system c, the expression of the target in the world coordinate system w is as follows:

[0048] (2)

[0049] wherein, represents the detection box corresponding to the i-th target in the camera coordinate system , , are the three-dimensional position and attitude of the i-th target in the camera coordinate system , is the length, width and height of the detection box corresponding to the i-th target.

[0050] According to the true pose of the target and the relative pose relationship between the target and the carrier, the pose of the camera in the world coordinate system w can be determined. At this time and have the following conversion relationship:

[0051] (3)

[0052] in, Represents the rotation matrix from the camera coordinate system c to the world coordinate system w.

[0053] Specifically, in step 2, the uniform acceleration and uniform angular velocity of six degrees of freedom are used to describe the target motion state. Assume that Target being tracked at all times The projection of the acceleration and angular velocity in the world coordinate system w in the target coordinate system is: and , then you can State prediction at each moment Status at the moment:

[0054] (4)

[0055] in, express Time by The rotation matrix from the target coordinate system to the world coordinate system w, 、 、 Indicates the target exist Position, velocity, and posture in the world coordinate system w at the moment; 、 、 Indicates the target exist Position, velocity, and posture in the world coordinate system w at the moment; Indicates the interval between two adjacent detection frames.

[0056] Specifically, in getting After the state is predicted at each moment, the target prediction trajectory is associated with the 3D detection bounding box. In order to determine the credibility of the association between the detected target and the tracked target, it is necessary to construct the association loss matrix C of each detected target and each tracked target. The elements in the matrix are is the association loss between the i-th detected target and the j-th tracked target. The higher the association loss between the detected target and the tracked target, the lower the association credibility between the two. The specific construction of the association loss is as follows:

[0057] (5)

[0058] in, is the intersection-over-union ratio of the 3D detection boxes corresponding to the i-th detected target and the j-th tracked target.

[0059] After the association loss matrix is ​​constructed, the association problem of the detection box can be solved by the Hungarian algorithm, and its best matching matrix for

[0060] (6)

[0061] in, represents the association loss between the i-th detected target and the j-th tracked target, Indicates whether the i-th detected target matches the j-th tracked target. The following conditions must be met at the same time:

[0062] (7)

[0063] At the same time, set a threshold for the association loss , when it is higher than the threshold, it is considered unassociatable to prevent false association.

[0064] Specifically, in order to improve the continuity of target tracking, target lifecycle management is used to assist data association. The time the target is tracked The existence and extinction of the maintenance target, the corresponding thresholds are and In an observation frame, if the tracked target matches the detected target successfully, Clear and update ;like Then optimize the state of the target and predict the state of the tracked target in the next observation frame; if Then directly update the target trajectory and make state predictions; if the tracked target fails to match all detected targets, it is considered that the tracked target disappears in this frame and the state prediction is made. To record the target disappearance time; if It is considered that the tracked target is in a temporary disappearance state. At this time, the state of the target predicted by the previous observation frame can be used as the state of this frame, and the state prediction can continue; if If a detected target in a detection frame does not successfully match any existing tracked target, the detected target is recorded as a new tracked target. Through the above process, the target life cycle management can be achieved to improve tracking accuracy.

[0065] Specifically, in step 2, the target state includes the position of the target in the world coordinate system ,speed ,attitude ; The projection of the acceleration and angular velocity of the target in the world coordinate system w in the target coordinate system 、 ; The size of the target detection box Therefore, the estimated state of the target motion within the sliding window can be expressed as:

[0066] (8)

[0067] in, 、 , is the state of the i-th target, For the i-th target The state at the moment, n is the number of targets, and L is the sliding window length.

[0068] Specifically, the target motion state optimization problem is actually a maximum posterior estimation problem. A cost function is constructed to solve it. The expression is as follows:

[0069] (9)

[0070] in, is the Cauchy robust kernel function, represents the prior information obtained by marginalization, is the target motion prediction trajectory residual, is the target 3D detection bounding box residual, is the zero-velocity constraint residual, is the non-holonomic constraint residual, O is the set of all targets, M is the set of target motion model measurements, is the set of all stationary targets, is the set of stationary target motion model measurements, and B is the set of target 3D detection bounding box measurements.

[0071] The cost function of formula (9) can be simplified as:

[0072] (10)

[0073] in, for arrive The system status to be estimated at the moment; is the covariance matrix of observations; for The observation model error at each moment. Equation (10) cannot be solved directly, so the observation equation exist The first-order Taylor expansion nearby gives:

[0074] (11)

[0075] in, is the observation equation About the pending status The first derivative of , i.e. the Jacobian matrix; is the increment of the state to be estimated. ,make reaches a minimum value, that is:

[0076] (12)

[0077] make ,get:

[0078] (13)

[0079] The function on the right side of equation (13) is about When the derivative of is zero, the minimum value can be obtained, that is:

[0080] (14)

[0081] The increment can be solved iteratively by equations (13) and (14): , and obtain the global optimal solution .

[0082] Specifically, if Figure 2 As shown in the figure, as the number of observation frames and target state dimensions in the factor graph increases, computational efficiency decreases. To avoid this problem, a sliding window approach is used to limit the number of observation frames involved in the optimization, achieving a balance between efficiency and accuracy. Within the sliding window, each optimization is performed only on the target state in the observation frames within the current window. At the end of the optimization, the sliding window will add and remove observation frames. The removed observation frames are marginalized and continue to influence the subsequent optimization process as prior information.

[0083] Specifically, for the nonlinear least squares problem, there is an incremental equation of the following form, which corresponds to equation (14) to obtain equation (15):

[0084] (15)

[0085] in, correspond , correspond , correspond .

[0086] After completing a nonlinear optimization, the state Can be divided into states that need to be removed The state that needs to be preserved , then formula (15) can be rewritten as:

[0087] (16)

[0088] Multiply the above formula by a matrix on the left to get:

[0089] (17)

[0090] Further calculation yields:

[0091] (18)

[0092] The above process is Schur's elimination. is the Schur complement matrix of H, removing the state from the above formula get:

[0093] (19)

[0094] At this time, the status Removed, but it affects the status The constraint still exists, so the above formula can be used as a priori constraint in the optimization process.

[0095] Specifically, the residuals in step 3 are constructed as follows:

[0096] (1) Zero-speed constraint is to use the physical characteristics of the carrier when it is stationary to constrain the parameters of the system to be estimated. This constraint plays an important role in suppressing the error divergence of the vehicle-mounted inertial navigation system. In the embodiment of the present invention, it is introduced into the target vehicle to derive the zero-speed constraint residual of the target vehicle. In the actual scenario, if the target The velocity data of the line with a velocity close to zero in the time period T will be closely distributed around the zero value. Therefore, the standard deviation of these velocity data is calculated by , we can get a quantitative indicator to evaluate the discreteness of the data and then determine whether the speed of the object is close to zero. The calculation formula is

[0097] (20)

[0098] in, is the number of observation frames in time period T, is the mean speed in time period T.

[0099] is the standard deviation Setting the threshold , if the calculated standard deviation Less than , then it is considered that the target For a stationary target, its theoretical pose remains unchanged at each moment during the stationary period, and the residual term can be constructed by comparing the pose differences detected at different moments during the stationary period of the target. and Represent the target At two different moments during the stationary period, the residual can be constructed according to the zero-speed constraint principle:

[0100] (twenty one)

[0101] in, is the corresponding vector part of the quaternion.

[0102] (2) The non-holonomic constraint refers to the assumption that the vehicle does not skid, drift, bounce, etc. during driving, and the lateral and vertical velocities of the vehicle are zero. This constraint is commonly found in systems such as wheeled mobile robots and vehicle-mounted integrated navigation. This paper extends it to the target vehicle and derives the non-holonomic constraint residual of the target vehicle. According to the definition, we get:

[0103] (twenty two)

[0104] in, for Time target The velocity component in the y-axis direction in the target coordinate system. The other parameters are similar. The directions of the axes in the target coordinate system are: the positive direction of the x-axis points to the target direction of travel, the positive direction of the y-axis points to the left of the target direction of travel, and the positive direction of the z-axis points to the zenith. The residual can be constructed according to the principle of non-holonomic constraints:

[0105] (twenty three)

[0106] in, For the vector The component corresponding to the axis.

[0107] Target by Predicted at any moment The state of the moment and its There are differences in the states detected at each moment, which is the target motion model residual. According to formula (4), it is defined as follows:

[0108] (twenty four)

[0109] Target exist The 3D bounding box detected at each moment can be expressed as Multi-target tracking systems usually assume that the camera poses It is known that at this time, the target state can be projected into the camera coordinate system and compared with the target 3D detection bounding box. The difference between the two is the target 3D detection bounding box residual, which is defined as follows:

[0110] (25)

[0111] The relevant parameters are as follows:

[0112] (26)

[0113] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functionality. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, on the basis of the above-mentioned embodiments, an embodiment of the present invention provides a 3D visual multi-target tracking system based on vehicle motion constraints. This system is used to implement a 3D visual multi-target tracking method based on vehicle motion constraints described in the above-mentioned method embodiment.

[0114] The system includes: a trajectory prediction module, which is used to predict the target object using a six-degree-of-freedom motion model to obtain the target motion prediction trajectory; a data association module, which is used to construct the data association between the target motion prediction trajectory and the target object, solve the data association using the Hungarian algorithm, and use target lifecycle management to assist the data association; an optimization model construction module, which is used to fuse the target motion prediction trajectory, the target object, the vehicle motion constraints and the prior information through a sliding window to construct a factor graph optimization model based on the sliding window; a target tracking module, which is used to optimize the motion state of the target object using a factor graph optimization model based on the sliding window, and input the optimized result into the six-degree-of-freedom motion model, and finally output the motion trajectory of the target object.

[0115] The 3D visual multi-target tracking system based on vehicle motion constraints provided by an embodiment of the present invention addresses the low accuracy of target state estimation and trajectory prediction in existing methods. By adopting several modules, it establishes a six-degree-of-freedom motion model for data association, thereby reducing target confusion and erroneous associations. By constructing a sliding window-based factor graph optimization model, it achieves joint and accurate estimation of the motion states of multiple targets, thereby improving the trajectory prediction and tracking accuracy of multiple targets.

[0116] Based on the same inventive concept as the above-mentioned embodiment, an embodiment of the present invention also provides an electronic device, including a memory and a processor, the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement a 3D visual multi-target tracking method based on vehicle motion constraints as proposed in the above-mentioned embodiment.

[0117] An embodiment of the present invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, this program overcomes the low accuracy of target state estimation and trajectory prediction in existing methods, enhances pose estimation accuracy and continuity in complex environments, and achieves robust 3D multi-target tracking performance.

[0118] The storage medium can be a hard disk, a solid state disk, a flash disk, an optical disk or any non-volatile storage device for storing computer program codes and necessary data files, and the stored computer programs include a trajectory prediction module, a data association module, an optimization model construction module and a target tracking module.

[0119] Finally, it should be pointed out that the above specific embodiments are only representative examples of the present application. Obviously, the present application is not limited to the above specific embodiments, and there can be many variations. Any simple modification, equivalent change and modification made to the above specific embodiments in accordance with the technical essence of the present application shall be considered to fall within the protection scope of the present application.

Claims

1. A 3D visual multi-target tracking method based on vehicle motion constraints, characterized in that: include: S1. Use the six-degree-of-freedom motion model to predict the target object and obtain the target motion prediction trajectory; S2. Construct data association between target motion prediction trajectory and target object, use Hungarian algorithm to solve data association, and use target lifecycle management to assist data association; S3, through the sliding window, the target motion prediction trajectory, target object, vehicle motion constraints and prior information are integrated to construct a sliding window-based factor graph optimization model; the vehicle motion constraints in S3 include zero-speed constraints and non-holonomic constraints: the zero-speed constraint refers to the use of the physical characteristics of the vehicle when it is stationary to constrain the parameters to be estimated, and the zero-speed constraint residual is constructed in combination with the target posture difference; the non-holonomic constraint refers to the assumption that the vehicle does not skid, drift, or bounce during the vehicle's driving process, and the lateral and vertical velocities of the vehicle are zero, and the non-holonomic constraint residual is constructed in combination with the orientation of each axis of the target coordinate system; S4. Optimize the motion state of the target object using a sliding window-based factor graph optimization model, input the optimized result into a six-degree-of-freedom motion model, and finally output the motion trajectory of the target object; the expression for optimizing the motion state of the target object in S4 is: ， in, is the Cauchy robust kernel function, is the prior information obtained by marginalization, is the target motion prediction trajectory residual, is the target 3D detection bounding box residual, is the zero-velocity constraint residual, is the non-holonomic constraint residual, O is the set of all targets, M is the set of target motion model measurements, is the set of all stationary targets, is the set of stationary target motion model measurements, and B is the set of target 3D detection bounding box measurements.

2. The 3D visual multi-target tracking method based on vehicle motion constraints according to claim 1 is characterized in that: Said S2 further includes: Represent the target object as a target 3D detection bounding box; Based on the target 3D detection bounding box and the target prediction trajectory, an association loss matrix is ​​constructed, and the Hungarian algorithm is used to solve the association loss matrix to obtain matching pairs.

3. The 3D visual multi-target tracking method based on vehicle motion constraints according to claim 1 is characterized in that: After optimizing the motion state of the target object by using the sliding window-based factor graph optimization model in S4, the method further includes: marginalizing the observation frames removed after the optimization as prior information.

4. The 3D visual multi-target tracking method based on vehicle motion constraints according to claim 2 is characterized in that: After getting the matching pairs, it also includes: Based on the historical motion trajectory of failed matching, a six-degree-of-freedom motion model is used to predict the state and match again; Based on the successfully matched matching pairs, they are input into the sliding window-based factor graph optimization model for motion state optimization; Based on the detected target that fails to match, the detected target is marked as a new target.

5. The 3D visual multi-target tracking method based on vehicle motion constraints according to claim 1 is characterized in that: The six-degree-of-freedom motion model in S1 includes six-degree-of-freedom constant acceleration and angular velocity, and is used to predict the motion trajectory of the target object.

6. A 3D visual multi-target tracking system based on vehicle motion constraints, applied to the 3D visual multi-target tracking method based on vehicle motion constraints according to claim 1, characterized in that: include: The trajectory prediction module is used to predict the target object using a six-degree-of-freedom motion model to obtain the target motion prediction trajectory; The data association module is used to construct the data association between the target motion prediction trajectory and the target object, using the Hungarian algorithm to solve the data association and using the target lifecycle management to assist the data association; The optimization model construction module is used to fuse the target motion prediction trajectory, target object, vehicle motion constraints and prior information through a sliding window to build a sliding window-based factor graph optimization model; The target tracking module is used to optimize the motion state of the target object using a factor graph optimization model based on a sliding window, input the optimized results into a six-degree-of-freedom motion model, and finally output the motion trajectory of the target object.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the 3D visual multi-target tracking method based on vehicle motion constraints according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the 3D visual multi-target tracking method based on vehicle motion constraints according to any one of claims 1 to 5 are implemented.

Images