Laser radar multi-target tracking system and method in complex scene

Abstract

The invention discloses a laser radar multi-target tracking system and method in a complex scene, the system comprises a detection module, a state prediction module, a data association module, a state updating module and a trajectory management module, a three-dimensional laser radar obtains original point cloud data in real time, and the detection module obtains a target detection result; performing prediction calculation on the position of each target in the later time period at the t-1 moment through a state prediction module; correlation matching is performed on the detection result and the prediction calculation result through a correlation module; an updating module is used for updating the prediction calculation result by taking the detection result which is successfully associated and matched in the step S3 as new observation data; track management is carried out on an associated result; in conclusion, the three-dimensional laser radar multi-target tracking method overcomes the defects in the prior art, is reasonable in design, provides an efficient and accurate three-dimensional laser radar multi-target tracking method in a complex scene, and has high social use value and application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of linear multi-target tracking, and particularly to a lidar multi-target tracking system and method in complex scenarios. Background Art

[0002] In recent years, with the complex and changeable application environment, it is required that the radar has the ability to track multiple targets and can simultaneously achieve multi-target tracking.

[0003] Existing methods are prone to mis-association or trajectory loss under occlusion, and it is difficult to effectively maintain the continuity of target IDs; most methods adopt a single-stage matching strategy and do not fully utilize features at different levels for multi-level matching, resulting in a decrease in association accuracy in complex scenarios. To solve the above problems, the present invention proposes an efficient lidar multi-target tracking system and method in complex scenarios to enhance the robustness, occlusion recovery ability, and real-time performance of target tracking. Summary of the Invention

[0004] To solve the problems mentioned in the above background art, the present invention provides a lidar multi-target tracking system and method in complex scenarios.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A lidar multi-target tracking system in complex scenarios, including a detection module, a state prediction module, a data association module, a state update module, and a trajectory management module;

[0007] The detection module is used to detect the original point cloud data obtained by the 3D lidar at the current time t, and obtain the detection result D of the target in the frame at time t t ;

[0008] The state prediction module performs prediction calculations on the positions of each target in the subsequent time period at time t - 1, and thus can obtain the predicted state T of the target in the frame at time t from the frame at time t - 1 est ;

[0009] The data association module associates and matches the detection result D of the frame at time t t and the predicted state T of the target in the frame at time t from the frame at time t - 1 est and outputs the detection result D with successful matching macth , the trajectory state T with successful matching macth and the detection result D that is not matched unmacth , the trajectory state T that is not matched unmacth ;

[0010] The state update module uses the detection result D with successful matching in the data association module macthAs new observation data, for each successfully matched track state T in data association macth is updated, and finally all tracks T at time t are obtained t ;

[0011] The track management module will use a certain time t min The unmatched detection results D that exist in consecutive frames within and have no matching tracks unmacth are initialized as new track states T new , and are incorporated into the track set; the track management module will use a long time t max Track states T that have not been matched by the detection results D t within are removed from the track set unmacth .

[0012] A multi-object tracking method for lidar in complex scenarios includes the following steps:

[0013] S1: The 3D lidar obtains raw point cloud data in real time and the detection module obtains target detection results;

[0014] S2: The state prediction module performs prediction calculations on the positions of each target in the subsequent time period at time t-1;

[0015] S3: The association module associates and matches the detection results with the prediction calculation results;

[0016] S4: The update module uses the detection results successfully associated in S3 as new observation data to update the prediction calculation results;

[0017] S5: Perform track management on the associated results.

[0018] Preferably, the specific steps of S2 are as follows: In order to accurately associate the candidate detection boxes with the existing tracks, for each tracking object at time t-1, it is necessary to predict the state at the current time t. For each target at each moment, an eleven-dimensional data set is used to represent its state information:

[0019] T = (x, y, z, θ, l, w, h, s, v x , v y , v z )

[0020] where x, y, and z are the centroid coordinates of the target in the three-dimensional coordinate system, θ is the heading angle of the target, l, w, and h are the three-dimensional dimensions of the target boundary, v x , v y , v zThey are the component velocities of the target in different directions in the three-dimensional coordinate system. Assuming there are k targets in the frame at time t-1, these k targets are represented by the following set:

[0021]

[0022] And a state transformation model is constructed based on the coordinates and component velocities of the target to estimate the pose of the target at time t:

[0023]

[0024] By predicting and calculating the positions of the targets in the time period after time t-1, the predicted states of all the targets in the frame at time t-1 in the current frame at time t can be obtained:

[0025] T est =(x est ,y est ,z est ,θ,l,w,h,s,v x ,v y ,v z ).

[0026] Preferably, the step S3 is specifically as follows:

[0027] S31: Adaptive Conversion Intersection over Union (AC-IoU):

[0028] First, a threshold σ is set. When the calculated value of the intersection over union is greater than σ, the intersection over union is reliable at this time, and this method is used for data association; when the value of the intersection over union is less than σ, it is unreliable to only use the intersection over union. At this time, the scale parameter ρ(A,B) and c are introduced as penalty terms; where ρ(A,B) represents the straight-line distance between the centroid of the 3D prediction box B and the ground truth box A, and c represents the diagonal length of the minimum bounding box of the 3D prediction box B and the ground truth box A; the converted intersection over union can directly minimize the distance between the two target boxes, and only two additional straight-line distance calculations are involved, with fast calculation speed and not much consumption of computing resources; the specific formula is as follows:

[0029]

[0030] S31: Cascade Matching:

[0031] The data association in the first stage uses the AC-IoU calculation method as the standard for successful matching, focusing on the degree of overlap between the associated tracking objects; first, the detection result D of the current frame is input t and the trajectory prediction result T of the targets at the previous moment at the current moment est, through a one-stage matching module, the cost matrix is calculated using the Adaptive Conversion Intersection over Union (AC-IoU), and then the cost matrix is fed into the Hungarian algorithm for matching; finally, the detection results D of successful one-stage matching are output macth1 , the trajectory status T of successful one-stage matching macth1 , and the detection results D of non-matched one-stage unmacth1 , the trajectory status T of non-matched one-stage unmacth1 ;

[0032] The second-stage data association is responsible for processing objects that cannot be successfully matched in the first stage; to support the successful re-association of unmatched objects, the second-stage association introduces the Euclidean distance and the heading angle to calculate the position correlation; the detection results D of non-matched one-stage unmacth1 , the trajectory status T of non-matched one-stage unmacth1 enter the second-stage matching module, and are further matched by combining the geometric position and the heading angle to solve the target matching problem in complex situations such as occlusion and low confidence, and output the detection results D of successful second-stage matching macth2 and the trajectory status T of successful second-stage matching macth2 , as well as the detection results D of finally non-matched unmacth2 and the trajectory status T of finally non-matched unmacth2 , and their cascaded matching is as Figure 2 shown;

[0033] Among them, the detection results D of successful matching macth include the detection results D of successful one-stage matching macth1 and the detection results D of successful second-stage matching macth2 , the trajectory status T of successful matching macth include the trajectory status T of successful one-stage matching macth1 and the trajectory status T of successful second-stage matching macth2 ; the detection results D of non-matching unmacth include the detection results D of finally non-matched unmacth2 , the trajectory status T of non-matching unmacth include the detection results D of finally non-matched unmacth2 .

[0034] Preferably, the step S4 is specifically as follows:

[0035] In Kalman filtering, state prediction is only based on the motion model. However, targets in the real environment may be affected by noise or various external factors, which can cause the prediction results to deviate significantly from the true values. In this chapter, by integrating the results of data association into the prediction results, the previous prediction data is reasonably corrected to obtain an estimated value closer to the true state. Considering that each target may undergo state mutations or offsets in the environment, single state prediction is unstable. Therefore, the detection results D macth matched successfully in data association are used as new observation data to update the state T macth of each trajectory matched successfully in data association, and finally all trajectories T t at time t are obtained:

[0036] T t ={T t 1 ,T t 2 ,T t 3 ,…,T t k}.

[0037] Preferably, the specific steps of step S5 are as follows: perform trajectory management on the associated results: when there are unmatched detection results D unmacth in the detection results, and when the unmatched detection results D unmacth exist in consecutive frames within a certain time t min and there are no matching trajectories, they will be initialized as new trajectories T new , a unique ID will be assigned to them and the state will be initialized through the Kalman filter, so as to incorporate them into the trajectory set for subsequent tracking to ensure that new targets can be captured in real time. For the unmatched trajectory states T max that have not been matched by the detection results within a long time t unmacth , the system will remove these trajectories from the trajectory set to clean up the invalid trajectories T lost , avoiding interference with subsequent tracking. The addition and deletion of trajectories work together, enabling the system to dynamically adapt to the appearance and disappearance of targets, thereby maintaining the rationality and efficiency of the trajectory set and improving the tracking performance and computational efficiency.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: In complex scenarios, the state prediction module can accurately predict the current state of each target. The AC-IoU used in the association stage of this method comprehensively considers the appearance and geometric features of the target, making the association more robust. Even in the case of target occlusion, complex motion, or long target spacing, it can effectively reduce false matches and missed matches. At the same time, the cascaded matching strategy further optimizes the accuracy of the association by processing trajectories and detection results with different priorities in stages, effectively eliminating the identity switching problem caused by occlusion or target overlap, especially in dealing with complex scenarios such as long-term occlusion or the initial appearance of new targets. In the trajectory management module, the addition and deletion of trajectories work together to enable the system to dynamically adapt to the appearance and disappearance of targets, thereby maintaining the rationality and efficiency of the trajectory set, and improving the tracking performance and computational efficiency.

[0039] In summary, the present invention overcomes the deficiencies of the prior art, is reasonably designed, and provides an efficient and accurate 3D lidar multi-target tracking method in complex scenarios, which has high social use value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0041] Figure 1 is a flowchart of the overall structure of the present invention;

[0042] Figure 2 is a flowchart of the cascaded matching of the present invention;

[0043] Figure 3 is a flowchart of the trajectory management of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0045] Embodiment 1

[0046] Refer to Figure 1, a multi-target tracking system for lidar in complex scenarios, including a detection module, a state prediction module, a data association module, a state update module, and a trajectory management module;

[0047] Further, the detection module is used to detect the original point cloud data obtained by the 3D lidar at the current time t, and obtain the detection result D of the target in the frame at time t t .

[0048] Further, the state prediction module performs prediction calculations on the positions of each target in the subsequent time period at time t-1, and the predicted state T of the target in the frame at time t-1 in the frame at time t can be obtained est , and the specific steps are as follows:

[0049] In order to accurately associate the candidate detection boxes with the existing trajectories, for each tracking object at time t-1, it is necessary to predict the state at the current time t. For each target at each moment, an eleven-dimensional data set is used to represent its state information:

[0050] T = (x, y, z, θ, l, w, h, s, v x , v y , v z )

[0051] Among them, x, y, and z are the centroid coordinates of the target in the three-dimensional coordinate system, θ is the heading angle of the target, l, w, and h are the three-dimensional dimensions of the target boundary, and v x , v y , v z are the component velocities of the target in different directions in the three-dimensional coordinate system. Assuming that there are k targets in the frame at time t-1, the following set is used to represent these k targets:

[0052]

[0053] And a state transformation model is constructed according to the coordinates and component velocities of the target to estimate the pose of the target at time t:

[0054]

[0055] By predicting and calculating the positions of the targets in the time period after time t-1, the predicted states of all targets in the frame at time t-1 in the current frame at time t can be obtained:

[0056] T est = (x est , y est , z est , θ, l, w, h, s, v x , v y , v z ).

[0057] The data association module associates the detection result D of the frame at time t t with the predicted state T of the target in the frame at time t-1 est for association and matching, and outputs the detection result D with successful matching macth , the trajectory state T with successful matching macth , and the detection result D that is not matched unmacth , and the trajectory state T that is not matched unmacth , and its specific principle is as follows:

[0058] 1. Adaptive Conversion Intersection over Union (AC-IoU):

[0059] First, a threshold σ is set. When the calculated value of the intersection over union is greater than σ, the intersection over union is reliable at this time, and this method is used for data association; when the value of the intersection over union is less than σ, it is unreliable to only use the intersection over union. At this time, the scale parameter ρ(A,B) and c are introduced as penalty terms. Among them, ρ(A,B) represents the straight-line distance between the centroid of the 3D prediction box B and the ground truth box A, and c represents the diagonal length of the minimum bounding box of the 3D prediction box B and the ground truth box A. The converted intersection over union can directly minimize the distance between the two target boxes, and only two additional straight-line distance calculations are involved. The calculation speed is fast and it does not consume too much computing resources. The specific formula is as follows:

[0060]

[0061] 2. Cascade matching:

[0062] The first-stage data association uses the AC-IoU calculation method as the criterion for successful matching, focusing on the overlap degree between the associated tracking objects. First, the detection result D of the current frame t and the trajectory prediction result T of the target at the previous moment at the current moment est are input. Through the first-stage matching module, the cost matrix is calculated using the adaptive conversion intersection over union (AC-IoU), and then the cost matrix is sent into the Hungarian algorithm for matching. Finally, the detection result D with successful first-stage matching macth1 , the trajectory state T with successful first-stage matching macth1 , and the detection result D that is not matched in the first stage unmacth1 , and the trajectory state T that is not matched in the first stage unmacth1 are output.

[0063] The second-stage data association is responsible for processing the objects that cannot be successfully matched in the first stage. To support the successful re-association of unmatched objects, the second-stage association introduces the Euclidean distance and the heading angle to calculate the position correlation. The detection result D that is not matched in the first stage unmacth1 , the trajectory state T that is not matched in the first stageunmacth1 Enter the second-stage matching module, further match by combining the geometric position and the heading angle, solve the target matching problem in complex situations such as occlusion and low confidence, and output the detection result D of successful second-stage matching macth2 and the trajectory state T of successful second-stage matching macth2 , as well as the finally unmatched detection result D unmacth2 and the finally unmatched trajectory state T unmacth2 , and their cascaded matching is as Figure 2 shown

[0064] Among them, the detection result D of successful matching macth includes the detection result D of successful first-stage matching macth1 and the detection result D of successful second-stage matching macth2 , the trajectory state T of successful matching macth includes the trajectory state T of successful first-stage matching macth1 and the trajectory state T of successful second-stage matching macth2 ; the unmatched detection result D unmacth includes the finally unmatched detection result D unmacth2 , the unmatched trajectory state T unmacth includes the finally unmatched detection result D unmacth2 .

[0065] Furthermore, the state update module uses the detection result D of successful matching in the data association module D macth as the new observation data to update each successfully matched trajectory state T in the data association macth , and finally obtain all the trajectories T at time t t , and its design steps are as follows

[0066] In Kalman filtering, state prediction is only based on the motion model, but the targets in the real environment may be affected by noise or various external factors, which will cause the prediction results to deviate significantly from the true values. In this chapter, by fusing the results of data association into the prediction results, the previous prediction data is reasonably corrected, so as to obtain an estimated value closer to the true state. Considering that each target may experience state mutations or offsets in the environment, a single state prediction is unstable. Therefore, the detection result D of successful matching in the data association macth is used as the new observation data to update each successfully matched trajectory state T in the data association macth , and finally obtain all the trajectories T at time t t :

[0067] T t ={T t 1 ,Tt 2 , T t 3 , …, T t k}。

[0068] Furthermore, the trajectory management module initializes the un - matched detection results D that exist in consecutive frames within a certain time t min and have no matching trajectories as new trajectory states T unmacth , and incorporates them into the trajectory set; the trajectory management module removes the un - matched trajectory states T that have not been matched by the detection results D new within a long time t max from the trajectory set. The specific design steps and principles are as follows: t unmacth unmacth The flowchart of trajectory management is as

[0069] shown in Figure 3 shown.

[0070] The multi - target tracking method for lidar in complex scenarios includes the following steps:

[0071] S1: The 3D lidar obtains the original point cloud data in real - time and the detection module gets the target detection results.

[0072] S2: The state prediction module performs prediction calculations on the positions of each target in the subsequent time period at time t - 1:

[0073] To accurately associate the candidate detection boxes with the existing trajectories, for each tracking object at time t - 1, it is necessary to predict the state at the current time t. For each target at each moment, an eleven - dimensional data set is used to represent its state information:

[0074] T = (x, y, z, θ, l, w, h, s, v x , v y , v z )

[0075] where x, y, and z are the centroid coordinates of the target in the three - dimensional coordinate system, θ is the heading angle of the target, l, w, and h are the three - dimensional dimensions of the target boundary, and v x , v y , v z are the component velocities of the target in different directions in the three - dimensional coordinate system. Assume that there are k targets in the frame at time t - 1, then these k targets are represented by the following set:

[0076]

[0077] And a state transformation model is constructed based on the coordinates and component velocities of the target to estimate the target pose at time t:

[0078]

[0079] By predicting and calculating the positions of the targets in the time period after time t - 1, the predicted states of all targets in the current time t frame in the time t - 1 frame can be obtained:

[0080] T est =(x est , y est , z est , θ, l, w, h, s, v x , v y , v z ).

[0081] S3: The detection results are associated and matched with the prediction calculation results through an association module:

[0082] S31: Adaptive Conversion Intersection over Union (AC-IoU):

[0083] First, a threshold σ is set. When the calculated value of the intersection over union is greater than σ, the intersection over union is reliable at this time, and this method is used for data association; when the value of the intersection over union is less than σ, it is unreliable to only use the intersection over union. At this time, the scale parameters ρ(A, B) and c are introduced as penalty terms. Among them, ρ(A, B) represents the straight-line distance between the centroid of the 3D prediction box B and the ground truth box A, and c represents the diagonal length of the minimum bounding box of the 3D prediction box B and the ground truth box A. The converted intersection over union can directly minimize the distance between the two target boxes, and only two additional straight-line distance calculations are involved. The calculation speed is fast and it does not consume too much computing resources. The specific formula is as follows:

[0084]

[0085] S31: Cascade matching

[0086] The one-stage data association uses the AC-IoU calculation method as the criterion for successful matching, focusing on the overlap degree between associated tracking objects. First, the detection results D t of the current frame and the trajectory prediction results T est of the targets at the previous moment in the current moment are input. Through the one-stage matching module, the cost matrix is calculated using the adaptive conversion intersection over union (AC-IoU), and then the cost matrix is sent into the Hungarian algorithm for matching. Finally, the detection results D macth1 that are successfully matched in the first stage, the trajectory states T macth1 that are successfully matched in the first stage, and the detection results D unmacth1, the trajectory state T not matched in the first stage unmacth1 .

[0087] The second-stage data association is responsible for processing the objects that cannot be successfully matched in the first stage. To support the successful re-association of unmatched objects, the second-stage association introduces the Euclidean distance and the heading angle to calculate the position correlation. The detection result D not matched in the first stage unmacth1 , the trajectory state T not matched in the first stage unmacth1 enter the second-stage matching module, and are further matched by combining the geometric position and the heading angle to solve the target matching problem in complex situations such as occlusion and low confidence, and output the detection result D successfully matched in the second stage macth2 and the trajectory state T successfully matched in the second stage macth2 , as well as the finally unmatched detection result D unmacth2 and the finally unmatched trajectory state T unmacth2 , and their cascaded matching is as shown in Figure 2 .

[0088] Among them, the successfully matched detection result D macth includes the detection result D successfully matched in the first stage macth1 and the detection result D successfully matched in the second stage macth2 , the successfully matched trajectory state T macth includes the trajectory state T successfully matched in the first stage macth1 and the trajectory state T successfully matched in the second stage macth2 ; the unmatched detection result D unmacth includes the finally unmatched detection result D unmacth2 , the unmatched trajectory state T unmacth includes the finally unmatched detection result D unmacth2 .

[0089] S4: Use the update module to update the prediction calculation result with the detection result successfully associated and matched in S3 as the new observation data:

[0090] In Kalman filtering, state prediction is only based on the motion model, but the targets in the real environment may be affected by noise or various external factors, which will cause the prediction result to deviate significantly from the true value. In this chapter, by integrating the result of data association into the prediction result, the previous prediction data is reasonably corrected, so as to obtain an estimated value closer to the true state. Considering that each target may undergo state mutation or deviation in the environment, a single state prediction is unstable. Therefore, the detection result D successfully matched in data association macth is used as the new observation data to update each successfully matched trajectory state T macth in data association, and finally all trajectories T t at time t are obtained:

[0091] T t = {T t 1 , T t 2 , T t 3 , …, T t k}}。

[0092] S5: Manage the trajectories for the associated results: When there are unmatched detection results D in the detection results unmacth , when the unmatched detection result D unmacth exists in consecutive frames within a certain time t min and there is no matching trajectory, it will be initialized as a new trajectory T new , a unique ID will be assigned to it and the state will be initialized through a Kalman filter, so as to incorporate it into the trajectory set for subsequent tracking, ensuring that new targets can be captured in real time. For the unmatched trajectory state T max that has not been matched by the detection results within a long time t unmacth , the system will remove these trajectories from the trajectory set to clean up the invalid trajectories T lost , avoiding interference with subsequent tracking. The addition and deletion of trajectories work together, enabling the system to dynamically adapt to the appearance and disappearance of targets, thus maintaining the rationality and efficiency of the trajectory set, and improving the tracking performance and computational efficiency.

[0093] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation of the present invention.

[0094] In the present invention, unless otherwise clearly specified and limited, the terms "set", "install", "connect", "couple", "fix", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium, and can be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0095] The control method of the present invention is automatically controlled by a controller. The control circuit of the controller can be realized by simple programming of those skilled in the art. The provision of power also belongs to the common general knowledge in this field. And the present invention is mainly used to protect mechanical devices. Therefore, the control method and circuit connection of the present invention will not be explained in detail herein.

[0096] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. The laser radar multi-target tracking system in complex scenes is characterized by: It includes detection module, state prediction module, data association module, state update module and trajectory management module; The detection module is used to detect the original point cloud data at the current time t acquired by the three-dimensional laser radar, and obtain the detection result D of the target at the time t frame t ; The state prediction module predicts the position of each target at time t-1 in the subsequent time period, and obtains the predicted state T of the target at time t in the frame at time t-1. est ; The data association module converts the detection result D of the frame at time t into t and the predicted state T of the target in the frame at time t-1 est Associate the matches and output the detection result D of successful matching macth , matching successful trajectory state T macth and unmatched test results D unmacth , unmatched trajectory state T unmacth ; The state update module updates the detection result D that is successfully matched in the data association module. macth As new observation data, each successfully matched trajectory state T in the data association macth Update and finally get all the trajectories T at time t t ; The trajectory management module sets a certain time t min Unmatched detection results D that exist in all consecutive frames and have no matching tracks unmacth Initialize to the new trajectory state T new , and incorporate it into the trajectory set; the trajectory management module will be long-term t max No test results were found in D t Matched and unmatched trajectory states T unmacth Remove the track collection.

2. A laser radar multi-target tracking method in a complex scene, used in the tracking system of claim 1, characterized in that: The following steps are involved: S1: The 3D laser radar acquires the original point cloud data in real time and the detection module obtains the target detection result; S2: The state prediction module is used to predict the position of each target at time t-1 in the subsequent time period; S3: Associating and matching the detection results with the prediction calculation results through the association module; S4: The prediction calculation result is updated by using the detection result of successful association matching in S3 as new observation data through the update module; S5: Perform track management on the associated results.

3. The laser radar multi-target tracking method in complex scenes according to claim 2, characterized in that: The step S2 is specifically as follows: In order to accurately associate the candidate detection box with the existing trajectory, for each tracked object at time t-1, it is necessary to predict the state at the current time t. For the target at each time, an eleven-dimensional data set is used to represent its state information: T=(x,y,z,θ,l,w,h,s,v x ,v y ,v z ) Among them, x, y, z are the centroid coordinates of the target in the three-dimensional coordinate system, θ is the heading angle of the target, l, w, h are the three-dimensional dimensions of the target boundary, and v x ,v y ,v z is the component velocity of the target in different directions in the three-dimensional coordinate system; assuming that there are k targets in the frame at time t-1, the following set is used to represent these k targets: And according to the target's coordinates and velocity components, a state transformation model is constructed to estimate the target's position at time t: By predicting the position of the target in the time period after time t-1, the predicted state of all targets in the time frame of time t-1 in the current time frame can be obtained: T est =(x est ,y est ,z est ,θ,l,w,h,s,v x ,v y ,v z )。 4. The laser radar multi-target tracking method in complex scenes according to claim 2, characterized in that: The step S3 is specifically as follows: S31: Adaptive Conversion Intersection over Union (AC-IoU): First, a threshold σ is set. When the calculated IoU value is greater than σ, the IoU is reliable and this method is used for data association. When the IoU value is less than σ, it is unreliable to use only the IoU. At this time, the scale parameters ρ(A, B) and c are introduced as penalty items. ρ(A, B) represents the straight-line distance between the centroid of the 3D prediction box B and the real box A, and c represents the diagonal length of the minimum bounding box of the 3D prediction box B and the real box A. The converted IoU can directly minimize the distance between the two target boxes, and only two straight-line distances are calculated. The calculation speed is fast and does not consume too much computing resources. The specific formula is as follows: S31: Cascade Matching: The first-stage data association uses the AC-IoU calculation method as the criterion for successful matching, focusing on the degree of overlap between associated tracked objects. First, the detection result D of the current frame is input. t And the trajectory prediction result T of the target at the previous moment at the current moment est , through the one-stage matching module, the adaptive conversion intersection over union (AC-IoU) is used to calculate the cost matrix, and then the cost matrix is ​​sent to the Hungarian algorithm for matching; finally, the detection result D of the successful one-stage matching is output macth1 , the trajectory state T of successful matching in the first stage macth1 , and the unmatched detection results D in the first stage unmacth1 , the unmatched trajectory state T in one stage unmacth1 ; The second stage of data association is responsible for processing objects that cannot be successfully matched in the first stage. In order to support the successful association of unmatched objects, the second stage association introduces Euclidean distance and heading angle to calculate the position correlation. The detection results of the unmatched objects in the first stage are D unmacth1 , the unmatched trajectory state T in one stage unmacth1 Enter the second-stage matching module, combine the geometric position and heading angle for further matching, solve the target matching problem in complex situations such as occlusion and low confidence, and output the detection result D of successful second-stage matching macth2 And the trajectory state T that successfully matches the second stage macth2 , and the final unmatched detection result D unmacth2 and the final unmatched trajectory state T unmacth2 , its cascade matching is shown in Figure 2; Among them, the detection result D macth Including the detection result D of successful matching in the first stage macth1 And the detection result D of successful matching in the second stage macth2 , matching successful trajectory state T macth Including the trajectory state T of the successful matching in the first stage macth1 And the trajectory state T that successfully matches the second stage macth2 ; Unmatched test results D unmacth Including the final unmatched detection results D unmacth2 , unmatched trajectory state T unmacth Including the final unmatched detection results D unmacth2 .

5. The laser radar multi-target tracking method in complex scenes according to claim 2, characterized in that: The step S4 is specifically as follows: In Kalman filtering, state prediction is based only on motion models, but targets in real environments may be affected by noise or various external factors, which may cause the prediction results to deviate significantly from the true values. This chapter integrates the results of data association into the prediction results and makes reasonable corrections to the previous prediction data to obtain an estimate closer to the true state. Considering that each target may experience a sudden state mutation or deviation in the environment, a single state prediction is unstable, so the detection result D that is successfully matched in data association is used as the prediction result. macth As new observation data, each successfully matched trajectory state T in the data association macth Update and finally get all the trajectories T at time t t : T t ={T t 1 ,T t 2 ,T t 3 ,…,T t k }。 6. The laser radar multi-target tracking method in complex scenes according to claim 2, characterized in that: The step S5 is specifically as follows: the track management of the associated results: when there is an unmatched detection result D in the detection result unmacth , when the unmatched detection result D unmacth In a certain period of time min When there is no matching trajectory in the continuous frames of new , assign a unique ID to it and initialize the state through the Kalman filter, so as to include it in the trajectory set for subsequent tracking, ensuring that the new target can be captured in real time; and for the long time t max There is no unmatched trajectory state T that is matched by the detection result unmacth , the system will remove these trajectories from the trajectory set to clean up invalid trajectories T lost , to avoid interference with subsequent tracking; the synergy of adding and deleting trajectories enables the system to dynamically adapt to the appearance and disappearance of targets, thereby maintaining the rationality and efficiency of the trajectory set and improving tracking performance and computational efficiency.

Images