Multi-target real-time tracking and trajectory prediction method and system under occlusion environment

Abstract

This invention discloses a method and system for real-time multi-target tracking and trajectory prediction under occlusion conditions, relating to the field of target tracking technology. The proposed scheme includes acquiring the current frame detection box and historical trajectory states, determining adjacent trajectories for each target based on these, generating predicted trajectory boxes for each target in the current frame based on the historical trajectory states, and forming a candidate association set based on the spatial proximity relationship between the predicted trajectory boxes and the current frame detection boxes. This application constructs an occlusion phase fingerprint jointly characterized by the dominant boundary, continuation direction, and boundary clipping residual sequence through cross-frame evolution analysis based on the inward shrinkage of each boundary of the detection box relative to the predicted trajectory box. This solves the problem in existing technologies that treat occlusion as merely random noise and struggle to identify structurally distorted observations, enabling the distinction between observations suitable for effective updating and those affected by directional occlusion and unsuitable for direct updating.

Description

Technical Field

[0001] This invention relates to the field of target tracking technology, and more specifically, this application relates to a method and system for real-time tracking and trajectory prediction of multiple targets in occluded environments. Background Technology

[0002] With the development of applications such as video perception, intelligent security, traffic analysis, and robot vision, systems often need to simultaneously identify and continuously track multiple moving targets in continuous images, and provide subsequent positional change results under online processing conditions. These tasks typically require maintaining the continuity of target identity and trajectory based on current observations and historical motion information without relying on future frames. Therefore, real-time multi-target tracking and trajectory prediction have become fundamental capabilities in related fields.

[0003] In existing solutions, target detection results are typically obtained from the current image first. Then, the current position of the target is estimated by combining historical trajectories, and association is completed according to position proximity, appearance similarity, or general constraints. Subsequently, the trajectory state is updated using the matched detection results. For occlusion scenarios, existing technologies often mitigate performance degradation by broadening the candidate range, retaining low-confidence results, or enhancing motion estimation capabilities.

[0004] However, in common scenarios where targets intersect, travel in parallel, or partially overlap, occlusion does not merely manifest as increased random noise, but rather causes a persistent structural loss in the observation area. This leads to the detection results deviating from the true target state in both the center position and scale representation. If such distorted observations are directly used for trajectory updates, the bias will accumulate over time and propagate to subsequent position estimations. Furthermore, the previously stable local spatial relationships between neighboring targets will be gradually disrupted, making subsequent matching more prone to confusion among multiple similar candidates. This makes it difficult for existing technologies to distinguish between observations that can be used for effective updates and those that are affected by occlusion and are not suitable for direct updates, ultimately leading to state drift, identity switching, and prediction instability. Therefore, a multi-target real-time tracking and trajectory prediction method and system under occlusion conditions are proposed to address this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this technical solution provides a method and system for real-time multi-target tracking and trajectory prediction in occluded environments, resolving the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this application provides a method for real-time multi-target tracking and trajectory prediction in occluded environments, the method comprising:

[0008] Obtain the current frame detection box and historical trajectory status, and determine the adjacent trajectories of each target accordingly;

[0009] Based on the historical trajectory status, a predicted trajectory box for each target in the current frame is generated, and a candidate association set is formed based on the spatial proximity relationship between the predicted trajectory box and the detection box in the current frame.

[0010] For each candidate pair in the candidate association set, the shrinkage of the detection box relative to each boundary of the predicted trajectory box in the candidate pair is extracted, and the dominant boundary and its continuation direction are determined by combining the shrinkage in the corresponding consecutive frames of the historical trajectory state. The boundary clipping residual sequence is constructed to determine the occlusion phase fingerprint of the corresponding target.

[0011] When the occlusion phase fingerprint indicates that the target has entered directional occlusion, the occlusion tangent is constructed along the dominant boundary continuation direction. The relative position order of the target and the adjacent trajectory in the occlusion tangent is recorded as a local sequential signature. In subsequent frames, candidate pairs consistent with the local sequential signature are retained to obtain a sequentially feasible candidate set.

[0012] Based on the dominant boundary continuation direction determined by the boundary clipping residual sequence, the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set are reconstructed by reverse geometric compensation to obtain the complete pseudo observation box of the corresponding target, and the observation uncertainty along the direction perpendicular to the occlusion tangent is determined.

[0013] The historical trajectory states of the targets corresponding to the updated sequential feasible candidate set are obtained by using the complete pseudo observation frame and observation uncertainty, and the predicted trajectory frame of the current frame is retained for targets that have not entered the sequential feasible candidate set. The updated trajectory state of each target in the current frame is obtained, and the predicted trajectory frame of the next frame is generated according to the updated trajectory state of the current frame.

[0014] Secondly, this application provides a multi-target real-time tracking and trajectory prediction system under occlusion conditions, used to implement the aforementioned multi-target real-time tracking and trajectory prediction method under occlusion conditions, including:

[0015] The target trajectory acquisition module is used to acquire the detection box and historical trajectory status of the current frame, and determine the adjacent trajectories of each target accordingly.

[0016] The candidate association set generation module is used to generate the predicted trajectory boxes of each target in the current frame based on the historical trajectory status, and to form a candidate association set based on the spatial proximity relationship between the predicted trajectory boxes and the detection boxes in the current frame.

[0017] The phase fingerprint construction module is used to extract the indentation of the detection box relative to each boundary of the predicted trajectory box in each candidate pair in the candidate association set, and combine the indentation in the corresponding continuous frames of the historical trajectory state to determine the dominant boundary and its continuation direction, and construct the boundary clipping residual sequence to determine the occlusion phase fingerprint of the corresponding target.

[0018] The sequential signature filtering module is used to construct the occlusion tangent along the dominant boundary continuation direction when the occlusion phase fingerprint indicates that the target has entered directional occlusion. It records the relative position order of the target and the adjacent trajectory in the occlusion tangent as a local sequential signature, and retains candidate pairs that are consistent with the local sequential signature in subsequent frames to obtain a sequentially feasible candidate set.

[0019] The detection box compensation and reconstruction module is used to perform reverse geometric compensation and reconstruction on the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set according to the dominant boundary continuation direction determined by the boundary clipping residual sequence, to obtain the complete pseudo observation box of the corresponding target, and to determine the observation uncertainty along the direction perpendicular to the occlusion tangent.

[0020] The trajectory state update module is used to update the historical trajectory state of the target corresponding to the sequential feasible candidate set using the complete pseudo observation frame and observation uncertainty, and to retain the predicted trajectory frame of the current frame for targets that have not entered the sequential feasible candidate set, so as to obtain the updated trajectory state of each target in the current frame, and generate the predicted trajectory frame of the next frame based on the updated trajectory state of the current frame.

[0021] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for real-time tracking and trajectory prediction of multiple targets under occlusion conditions.

[0022] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for real-time multi-target tracking and trajectory prediction under occlusion conditions.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This application constructs an occlusion phase fingerprint by cross-frame evolution analysis based on the shrinkage of each boundary of the detection box relative to the predicted trajectory box. This fingerprint is jointly characterized by the dominant boundary, the continuation direction, and the boundary clipping residual sequence. This solves the problem that existing technologies treat occlusion as random noise and make it difficult to identify structurally distorted observations. It enables the distinction between observations that can be effectively updated and those that are affected by directional occlusion and are not suitable for direct updates. This suppresses the error accumulation and prediction instability caused by distorted observations being written into the trajectory state from the source.

[0025] This application solves the problem of confusion of neighboring candidates caused by existing technologies relying only on position proximity, appearance similarity or general constraints in target intersection, parallel passage and local overlap scenarios. It extracts the relative position order of the target and adjacent trajectories along the tangential direction of the occlusion after the target enters the directional occlusion, and solidifies it into a local sequential signature to filter the sequential feasible candidate set. It realizes the association constraint that can still maintain the continuity of local topological relationship under the conditions of center offset and appearance disturbance, and reduces the risk of identity switching and mismatch.

[0026] This application achieves complete pseudo-observation boxes by performing reverse geometric compensation reconstruction on the occluded detection boxes based on the dominant boundary continuation direction. It also implements selective state updates by combining the observation uncertainty along the direction perpendicular to the occlusion tangent. This solves the state drift problem caused by directly using the distorted detection results in the center position and width and height scales for trajectory updates in the prior art. It realizes the directional repair and limited utilization of occluded observations, suppresses the error correction in the occlusion direction while retaining effective observation information, and improves the continuity and stability of real-time tracking and trajectory prediction of multiple targets in occluded environments. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0028] Figure 1 The flowchart shows the multi-target real-time tracking and trajectory prediction method under occlusion environment proposed in this invention.

[0029] Figure 2 This is a structural block diagram of the multi-target real-time tracking and trajectory prediction system under occlusion conditions proposed in this invention. Detailed Implementation

[0030] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0031] Reference Figure 1 As shown, this application proposes a method for real-time multi-target tracking and trajectory prediction in occluded environments, including:

[0032] Obtain the current frame detection box and historical trajectory status, and determine the adjacent trajectories of each target accordingly;

[0033] It should be noted that the detection bounding box of the current frame can be directly output after the target detection processing of the current video frame; the historical trajectory state can adopt the target state results after the update of the previous frame; the adjacent trajectory is preferably other trajectories that have a spatial proximity relationship with the target at the current moment and may affect the occlusion judgment and order constraints in the future. It can be determined based on the center distance, boundary proximity relationship or local overlap relationship between the predicted trajectory boxes corresponding to the historical trajectory state, so as to establish occlusion analysis and association constraints around the local target cluster rather than all targets in the world in the future.

[0034] Based on the historical trajectory status, a predicted trajectory box for each target in the current frame is generated, and a candidate association set is formed based on the spatial proximity relationship between the predicted trajectory box and the detection box in the current frame.

[0035] It should be noted that the predicted trajectory boxes can be extrapolated to the current frame from the position, velocity, and box size information in the historical trajectory states. Based on this, using each predicted trajectory box as a reference, only the detection boxes in the current frame that have a spatial proximity relationship with it are retained as candidate objects, thus forming a candidate association set. The purpose is to establish a local association range through the predicted trajectory boxes, and then continue to identify the structural clipping phenomenon caused by occlusion within the local association range, so that subsequent processing focuses on the candidate pairs that are most likely to correspond to the same target, avoiding the expansion of false associations caused by too many global candidates in the case of occlusion.

[0036] For each candidate pair in the candidate association set, the shrinkage of the detection box relative to each boundary of the predicted trajectory box in the candidate pair is extracted, and the dominant boundary and its continuation direction are determined by combining the shrinkage in the corresponding consecutive frames of the historical trajectory state. The boundary clipping residual sequence is constructed to determine the occlusion phase fingerprint of the corresponding target.

[0037] It should be noted that the shrinkage amount is not detection noise in the general sense, but rather a geometric amount used to characterize the continuous shrinkage of the detection box relative to the predicted trajectory box in a certain boundary direction; if a certain boundary shows a continuously increasing shrinkage trend in consecutive frames, it indicates that the target is more likely to be directionally occluded in the corresponding direction of the boundary.

[0038] Furthermore, by constructing the inward shrinkage change of the dominant boundary in consecutive frames into a boundary clipping residual sequence, the geometric distortion of a single frame can be transformed into a cross-frame evolution feature, thereby unifying the encoding of "occlusion occurrence" and "how occlusion continues to develop" into an occlusion phase fingerprint, which is different from ordinary random jitter or occasional detection offset.

[0039] When the occlusion phase fingerprint indicates that the target has entered directional occlusion, the occlusion tangent is constructed along the dominant boundary continuation direction. The relative position order of the target and the adjacent trajectory in the occlusion tangent is recorded as a local sequential signature. In subsequent frames, candidate pairs consistent with the local sequential signature are retained to obtain a sequentially feasible candidate set.

[0040] It should be noted that when directional occlusion is known to continue developing along a certain dominant boundary, the occlusion tangent perpendicular to it can more stably reflect the relative arrangement relationship between the target and its neighboring targets. Therefore, by recording the relative position order of the target and its adjacent trajectories in the occlusion tangent, a local order signature that is stable for short-term occlusion can be obtained. Subsequent frames only retain candidate pairs that still satisfy this local order signature. In essence, under the condition that "appearance is unreliable and the center position has shifted", a local topological constraint that does not depend on appearance features and is more robust to clipping phenomena is introduced, thereby reducing the probability of identity switching in intersection, parallel passage and local overlap scenes.

[0041] Based on the dominant boundary continuation direction determined by the boundary clipping residual sequence, the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set are reconstructed by reverse geometric compensation to obtain the complete pseudo observation box of the corresponding target, and the observation uncertainty along the direction perpendicular to the occlusion tangent is determined.

[0042] It should be noted that, based on the dominant boundary and its continuation direction, the geometric loss on the side to be clipped is compensated in reverse, thereby obtaining a complete pseudo-observation box that is closer to the true range of the target. On this basis, the observation uncertainty is determined separately along the direction perpendicular to the shading tangent, which is the main propagation direction of the shading effect. The purpose is that the geometric distortion caused by shading usually has obvious directionality. If the observation confidence of each direction is still used equally, the systematic deviation in the shading direction will be incorrectly injected into the trajectory state. By modeling the directional uncertainty, it is beneficial to suppress the erroneous update in the shading direction while retaining the effective observation information.

[0043] The historical trajectory state of the target corresponding to the updated sequential feasible candidate set is obtained by using the complete pseudo observation frame and observation uncertainty, and the predicted trajectory frame of the current frame is retained for the target that has not entered the sequential feasible candidate set. The updated trajectory state of each target in the current frame is obtained, and the predicted trajectory frame of the next frame is generated according to the updated trajectory state of the current frame.

[0044] It should be noted that candidate pairs are only allowed to be written into the trajectory state as the basis for updating when they simultaneously satisfy occlusion phase recognition, local order consistency, and geometric compensation availability. For targets that have not entered the sequentially feasible candidate set, the predicted trajectory box of the current frame is retained as the state expression of that frame, instead of forcibly using unreliable detection results to complete the update. This avoids the secondary pollution of the target state by distorted detection boxes in the occlusion environment and allows the prediction of the next frame to be based on a more stable trajectory state.

[0045] It should be noted that the acquisition of the current frame detection box can be achieved using object detection techniques, such as object detection networks based on convolutional neural networks or Transformer structures; the generation of predicted trajectory boxes from historical trajectory states can be achieved using linear motion extrapolation methods or Kalman filter prediction methods; and trajectory state updates themselves can also be achieved using a state estimation framework.

[0046] Through the above technical solution, this embodiment models the boundary shrinkage evolution of the detection box relative to the predicted trajectory box, and uses the resulting occlusion phase fingerprint, local sequential signature and complete pseudo observation box together for trajectory update. This enables the system to identify distorted observations that are not suitable for direct update in scenarios of target intersection, parallel passage and local overlap. At the same time, it performs directional compensation and uncertainty control on recoverable observations, thereby effectively suppressing state drift, identity switching and trajectory prediction instability.

[0047] In an optional embodiment, the dominant boundary and its continuation direction are determined by combining the indentation amount in consecutive frames corresponding to the historical trajectory state, and a boundary clipping residual sequence is constructed to determine the occlusion phase fingerprint of the corresponding target, specifically including:

[0048] Within a preset time window, extract continuous historical detection boxes and their corresponding continuous historical predicted trajectory boxes from the historical trajectory status.

[0049] It should be noted that the preset time window is used to limit the length of consecutive frames involved in occlusion determination, so that the dominant boundary determination is based on short-term continuous evolution rather than single-frame occasional offset. Its value range is preferably 3 to 7 frames, and a typical value can be 5 frames. If the preset time window is too short, it will be difficult to distinguish between random jitter and true directional occlusion. If the preset time window is too long, it will introduce outdated deformation information and increase online latency.

[0050] For example, let the current frame index be... The preset time window length is The extracted continuous historical detection boxes and continuous historical predicted trajectory boxes cover the following frame range: Together with the current frame, it constitutes the time range used to compute the shrinkage sequence within the boundary.

[0051] Extract the boundary shrinkage of the current frame detection box relative to the current frame predicted trajectory box, and extract the historical boundary shrinkage of each consecutive historical detection box relative to its corresponding consecutive historical predicted trajectory box, and construct the boundary shrinkage sequence of each boundary in chronological order.

[0052] For example, the indentation is constructed based on the coordinate differences between the predicted trajectory box and the detection box in the four boundary directions:

[0053] Let the predicted trajectory box of the k-th frame be... The detection box in the k-th frame is ,in, , These are the coordinates of the left boundary, top boundary, right boundary, and bottom boundary of the predicted trajectory box and the detection box, respectively.

[0054] The indentation of the four boundaries can be expressed as follows: , , , ,in, , , , These represent the indentation amounts of the left, right, top, and bottom boundaries, respectively; if the corresponding boundary does not shrink inwards, the indentation amount is recorded as 0.

[0055] Based on this, the shrinkage amounts of the same boundary within a preset time window can be arranged chronologically to obtain a boundary shrinkage sequence. ;

[0056] Extract monotonically increasing gradient features from the boundary shrinkage sequence of each boundary along the time dimension. Determine the boundary with the monotonically increasing gradient features exceeding the preset deformation threshold and the largest cumulative shrinkage increment as the dominant boundary. Determine the normal direction pointing into the detection box that continuously increases the shrinkage of the boundary corresponding to the dominant boundary as the continuation direction.

[0057] For example, a normalized positive increment is used to characterize the monotonically increasing gradient feature: let the boundary... In the The reference scale corresponding to the frame is ,when When it is the left or right boundary, The width of the predicted trajectory box can be selected; when When it is the upper or lower boundary, If we can take the height of the predicted trajectory box, then we have:

[0058] , ;

[0059] In the formula, For the boundary Normalized positive inward shrinkage gradient between two adjacent frames For this boundary The cumulative shrinkage increment within a preset time window;

[0060] and The largest boundary is determined as the dominant boundary: ,in, As the dominant boundary, A preset deformation threshold is set. The preset deformation threshold is set to distinguish between "continuous boundary clipping caused by the target being actually occluded by directional occlusion" and "accidental boundary shrinkage caused by detector jitter, slight pose changes, or box regression errors". The preferred value range is 0.05 to 0.30, and a typical value is 0.12. If the threshold is set too low, random detection noise will be misjudged as occlusion evolution. If the threshold is set too high, newly formed directional occlusion may be missed. The continuing direction is the normal of the dominant boundary pointing into the detection box, which represents the geometric direction in which the occlusion clipping continues to advance into the target.

[0061] Extract the boundary shrinkage of the dominant boundary in the current frame and each frame within the preset time window, and construct a boundary clipping residual sequence based on the difference in shrinkage between two adjacent frames in time.

[0062] For example, the dominant boundary The difference between adjacent frames is used as the boundary clipping residual sequence, expressed as:

[0063] ;

[0064] In the formula, This represents the change in the dominant boundary shrinkage between two adjacent frames. , The dominant boundary is at the th Frame and the The frame's indentation sequence within the frame boundaries; when When the value is continuously positive, it indicates that the occlusion boundary is continuously advancing inward; when... When fluctuations occur within a small range, it indicates that the occlusion is in a relatively stable stage. By constructing a residual sequence, the occlusion representation can simultaneously retain directional information and stage evolution information, rather than just the static cropping degree of a certain frame.

[0065] The dominant boundary, continuation direction, and boundary clipping residual sequence are established as the occlusion phase fingerprint of the corresponding target;

[0066] For example, the expression for occluding phase fingerprint is:

[0067] ;

[0068] In the formula, To obscure the phase fingerprint, As the dominant boundary, The normal vector corresponding to the direction of continuation. The boundary clipping residual sequence; the occlusion phase fingerprint makes the single boundary clipping phenomenon no longer just roughly represented by "a certain side is blocked", but further encoded as a composite feature of "which side is continuously clipped, in what direction the clipping progresses, and at what stage the clipping change is", which is conducive to subsequent linkage with local sequential signature and geometric compensation logic;

[0069] Through the above technical solution, this embodiment performs cross-frame serialization analysis on the four-boundary shrinkage of the detection box relative to the predicted trajectory box, and uses the dominant boundary, continuation direction and boundary clipping residual sequence to jointly construct the occlusion phase fingerprint, so that directional occlusion can be stably identified in a calculable and comparable form, thereby providing a consistent judgment basis for subsequent local sequence constraints and geometric compensation reconstruction.

[0070] In an optional embodiment, after extracting the monotonically increasing gradient features in the boundary shrinkage sequence of each boundary along the time dimension, the method further includes a non-formation determination process for cases where the monotonically increasing gradient features of each boundary do not reach a preset deformation threshold. Specifically, this includes:

[0071] Determine whether the monotonically increasing gradient features of the corresponding target at each boundary in the current frame have not reached the preset deformation threshold. If the determination is yes, record that the corresponding target is in the state of not reaching the threshold.

[0072] For example, when the condition is met When the target has not yet formed a significant boundary clipping trend sufficient to characterize directional occlusion within the preset time window, it can be determined that the target has not reached the threshold state. The purpose is not to simply output a binary result of "occlusion / non-occlusion", but to distinguish targets that have not yet formed clear directional clipping features from targets that have formed occlusion phase fingerprints, so as to avoid prematurely entering the dominant boundary determination and subsequent geometric compensation process.

[0073] In the state where the threshold is not reached, the historical boundary shrinkage corresponding to the earliest historical frame within the preset time window is moved out, and the boundary shrinkage of the current frame detection box relative to the current frame prediction trajectory box is moved in, so as to update the boundary shrinkage sequence of each boundary for the determination of the dominant boundary in the next frame.

[0074] It should be noted that this step can be implemented using a sliding window approach: when the current target has not yet formed a significant dominant boundary, the continuous observation of the boundary shrinkage is not terminated. Instead, the preset time window is slid forward one frame, allowing the latest observation to enter the sequence and the oldest observation to exit the sequence, so as to maintain a boundary shrinkage sequence of fixed length and temporal continuity. The purpose is that directional occlusion is often not formed instantaneously, but is gradually established as the target converges or overlaps. Therefore, using a sliding window to continuously accumulate the latest boundary information can enable subsequent frames to trigger the effective identification of the dominant boundary more quickly without increasing the additional storage burden.

[0075] Based on the state of not reaching the threshold, the occlusion phase fingerprint of the corresponding target is marked as not formed, and the process of determining the dominant boundary of the current frame is terminated.

[0076] It should be noted that when the current frame does not show a persistent clipping trend sufficient to support the establishment of the dominant boundary within the given observation window, the dominant boundary and the direction of continuation should not be forcibly output based on the current data. Instead, the occlusion phase fingerprint should be clearly marked as not formed so that the target is regarded as not entering the directional occlusion processing branch in the subsequent state update stage. This can avoid mistaking sporadic frame offsets as the initial phase of occlusion, thereby ensuring that the formation of the occlusion phase fingerprint is supported by continuous frame evidence, rather than originating from a single frame anomaly.

[0077] Through the above technical solution, this embodiment sets a state where the threshold is not reached and performs sliding updates on the boundary shrinkage sequence in this state. This prevents the system from prematurely outputting the dominant boundary and occlusion phase fingerprint when directional occlusion has not yet formed clear boundary evidence. This reduces misjudgments caused by detection jitter, slight pose changes, or local bounding box regression errors, and improves the stability and reliability of subsequent occlusion determination.

[0078] In an optional embodiment, the relative position order of the target and adjacent trajectories along the occlusion tangent is recorded as a local sequence signature. Candidate pairs consistent with this local sequence signature are retained in subsequent frames to obtain a sequence-feasible candidate set, specifically including:

[0079] Extract the center coordinates of the detection box in the current frame and the center coordinates of the predicted trajectory boxes in the current frame corresponding to each adjacent trajectory. Project the extracted center coordinates onto the occlusion tangent to obtain the corresponding one-dimensional projected coordinates.

[0080] For example, the local target relationships in a two-dimensional plane are reduced to a stable direction, the occlusion tangent, and expressed as follows: Let the center coordinates of the detection box in the current frame be... The center coordinates of the predicted trajectory boxes in the current frame corresponding to adjacent trajectories are: The unit vector of the occlusion tangential direction is Then its one-dimensional projected coordinates on the occlusion tangent can be expressed as: ;

[0081] in, Let be the projected coordinates of the center of the current frame detection box on the occlusion tangent. The projection coordinates of the centers of adjacent trajectories on the occlusion tangent are given. After adopting this one-dimensional projection, the local relative relationships that were originally greatly affected by the offset of the occlusion box center in two-dimensional space can be transformed into a sequential arrangement relationship along the tangent, which makes it easier to construct a sequence constraint that has stability for occlusion in the subsequent construction.

[0082] Based on the one-dimensional projection coordinates corresponding to the current frame detection box, determine the lateral distribution of each adjacent trajectory on the positive or negative side of the occlusion tangential direction, and establish the local order relationship of each side according to the distance between the one-dimensional projection coordinates of each adjacent trajectory and the current frame detection box from near to far.

[0083] For example, if If the corresponding adjacent trajectory is located on the positive side of the occlusion tangential direction; if If it is on the opposite side, then it is located on the reverse side; and the tangential distance between it and the current frame detection box is... It can be represented as Based on this, it can be pressed on both the forward and reverse sides respectively. The targets are sorted from smallest to largest to obtain the local order relationships on each side. The purpose is to extract the local topological relationships between the target and its neighboring targets in the tangential direction of occlusion after it has been determined that the target has entered directional occlusion, so as to form a more robust constraint basis for subsequent associations.

[0084] Extract the nearest neighboring trajectory identifier to the current frame detection box from the local order relationship on each side, and the distribution of the adjacent trajectory on the forward or reverse side, as a local order signature;

[0085] For example, the nearest neighbor relationship of the current detection box in the occlusion tangent direction is compressed into a concise and traceable local sequential signature. If the nearest adjacent trajectory is identified as... Then there is Furthermore, the local sequential signature can be represented as Among them, local sequential signature It also records "who is the nearest neighbor" and "which side the nearest neighbor is on". This signature method can reduce the storage burden of the local adjacency order on the one hand, and retain the local adjacency constraints that are most sensitive to identity confusion on the other hand.

[0086] In subsequent frames, the center coordinates of the candidate detection boxes of each candidate pair in the candidate association set are extracted, as well as the center coordinates of the predicted trajectory boxes corresponding to the adjacent trajectory identifiers in the local sequential signature. The extracted center coordinates are then projected onto the occlusion tangent to determine the candidate lateral distribution and candidate adjacency relationship of each candidate pair.

[0087] It should be noted that the same projection method as the current frame can be used to project the center of the candidate detection box in the subsequent frame and the center of the adjacent trajectory recorded in the local sequential signature to the same occlusion tangent, so as to obtain their relative lateral distribution and adjacency distance in the tangent. The purpose is that the local sequential signature does not directly depend on the appearance features of the detection box, but uses "which side of the adjacent trajectory the target should be located on in the occlusion tangent and which adjacent trajectory it should still be closest to" as the association verification condition for the subsequent frame. Therefore, it has better robustness to short-term occlusion, partial box deformation and appearance instability.

[0088] Candidates whose lateral distribution is consistent with that in the local sequential signature are retained, and whose candidate adjacency relationship indicates that the candidate detection box and the predicted trajectory box corresponding to the adjacent trajectory label maintain the nearest neighbor relationship, constitute the sequentially feasible candidate set;

[0089] It should be noted that a dual sequential screening is performed on the candidate association set: First, the lateral distribution of the candidate detection boxes in the occlusion tangent direction should be consistent with the local sequential signature; second, the adjacency relationship of the candidate detection boxes relative to the recorded adjacent trajectories is still the nearest neighbor; only when both conditions are met is the candidate pair retained as a sequentially feasible candidate set; this ensures that the local sequential constraint retains both directional information and minimum adjacency topology information, thereby avoiding identity swapping based solely on center proximity in multi-target intersections or parallel passages.

[0090] Through the above technical solution, this embodiment projects the center coordinates of the current frame detection box and adjacent trajectories to the occlusion tangent, and constructs a local sequential signature and a sequential feasible candidate set based on the projection results. This enables the subsequent association process to maintain the consistency constraint on the local topological relationship even when the center of the detection box is occluded and offset, thereby effectively reducing mismatches and identity switching in target intersection and local overlap scenarios.

[0091] In an optional embodiment, the center position and width / height dimensions of the detection boxes in the sequentially feasible candidate set are reconstructed using reverse geometric compensation to obtain the complete pseudo-observation box of the corresponding target, specifically including:

[0092] Extract the coordinates of the opposing boundary relative to the dominant boundary in the detection box of the feasible candidate set, as well as the coordinates of the two side boundaries intersecting with the dominant boundary;

[0093] It should be noted that once the dominant boundary has been determined by the occlusion phase fingerprint, the opposing boundary usually corresponds to the side that has not been truncated, while the two boundaries that intersect with the dominant boundary usually retain the currently observable side length range. Therefore, these three sets of boundary coordinates can be used as stable anchor points for reverse geometric compensation. The purpose is not to blindly enlarge the entire detection box, but to perform restricted reconstruction only around the dominant boundary that is most likely to be truncated, thereby reducing geometric distortion caused by unfounded expansion.

[0094] Extract three consecutive historical detection boxes from the historical trajectory state that correspond to the unoccluded phase fingerprint, and calculate their historical aspect ratio.

[0095] For example, it is preferable to select only the historical detection frames that have not formed occluded phase fingerprints as shape references to avoid introducing the frame size that has already been contaminated by occlusion and cropping back into the compensation process. Let the width and height of the three consecutive historical detection frames be respectively... , , The historical average width-to-height ratio can be expressed as: ;

[0096] in, This serves as a stable shape ratio reference for the current target under unobstructed or non-directionally occluded conditions. The purpose of using three consecutive frames instead of a single frame is to reduce the impact of instantaneous detection jitter on the ratio estimation. At the same time, it avoids using excessively long histories to prevent the ratio reference from becoming outdated due to changes in target pose or scale. This reflects the non-arbitrariness of using "short-term, unobstructed self-historical shape" as the compensation basis in this scheme.

[0097] Extract the current side length of the detection frame along the length of the dominant boundary, and combine it with the historical average width-to-height ratio to determine the compensation size along the continuity direction;

[0098] For example, a case-by-case calculation is performed based on the direction of the dominant boundary:

[0099] If the dominant boundary is the left or right boundary, then the length direction of the dominant boundary corresponds to the current height of the detection box. At this point, the historical average width-to-height ratio can be used. Restore the full width to obtain the compensated dimension along the continuity direction, i.e., the reconstructed width. If the dominant boundary is the upper or lower boundary, then the length direction of the dominant boundary corresponds to the current width of the detection box. At this point, the complete height can be restored, and the compensated dimensions along the continuity direction can be obtained. ;

[0100] in, The full width is obtained by reconstructing along the horizontal direction. To reconstruct the complete height along the vertical direction, and These are the observable height and observable width of the current detection box, respectively. The key to this compensation method is that instead of using a fixed prior box size, the continuous shape statistics of the target itself during the unoccluded phase are used to constrain the current reconstructed size, so that the compensation result is consistent with the target's historical true shape.

[0101] Keeping the coordinates of the opposing boundary and the coordinates of the two side boundaries fixed, using the coordinates of the opposing boundary as a reference, the coordinates of the occluded boundary are back-calculated according to the compensation size, and then combined with the coordinates of the opposing boundary and the coordinates of the two side boundaries to form a complete pseudo-observation frame.

[0102] For example, perform the corresponding reverse solution according to the different dominant boundaries:

[0103] When the dominant boundary is the left boundary, the occluded boundary is still the left boundary. The coordinates of the right boundary and the top and bottom boundaries can remain unchanged, and based on... The coordinates of the left boundary are obtained: ,in, The coordinates of the left boundary of the reconstructed complete pseudo-observation frame. To maintain a fixed right boundary coordinate;

[0104] Similarly, when the dominant boundary is the right boundary, the coordinates of the right boundary can be obtained: ,in, The coordinates of the right boundary of the reconstructed complete pseudo-observation frame. To maintain fixed left boundary coordinates;

[0105] When the dominant boundary is the upper boundary, the coordinates of the upper boundary can be obtained: ,in, The coordinates of the upper boundary of the reconstructed complete pseudo-observation frame. To maintain fixed lower boundary coordinates;

[0106] When the dominant boundary is the lower boundary, the coordinates of the lower boundary can be obtained: ,in, The coordinates of the lower boundary of the reconstructed complete pseudo-observation frame. To maintain fixed upper boundary coordinates;

[0107] To further explain, by recombining the reconstructed boundary with the original opposing boundary and lateral boundary, a complete pseudo-observation box can be obtained; the reconstruction process is "one-sided reverse compensation under boundary constraints", rather than uniform scaling of the entire box, so it is more suitable for handling local box missing under directional occlusion.

[0108] Through the above technical solution, this embodiment calculates the average historical width and height ratio using three consecutive historical detection frames that do not form an occlusion phase fingerprint, and uses the opposing boundary and the two side boundaries as fixed anchor points to perform reverse geometric compensation reconstruction on the occluded boundary. This allows the complete pseudo observation frame to not only restore the missing geometric range of the target in the occlusion direction, but also maintain consistency with the target's own historical shape, thereby providing a more reliable observation expression for subsequent state updates than the original occluded detection frame.

[0109] In an optional embodiment, determining the observation uncertainty along a direction perpendicular to the tangential direction of the occlusion specifically includes:

[0110] Extract the actual observed area of ​​the detection box and the reconstructed area of ​​the complete pseudo-observation box from the feasible candidate set of the order, and determine the area missing ratio accordingly;

[0111] For example, let the actual observation area of ​​the detection frame be... The reconstructed area of ​​the complete pseudo-observation frame is The area missing ratio can then be expressed as: ;in, The larger the value, the more area the current detection box loses relative to the reconstructed complete pseudo-observation box, indicating a more obvious degree of occlusion and clipping. It can reflect the incompleteness of the current observation from the perspective of geometric coverage, and is therefore suitable as the first basis for generating observation uncertainty.

[0112] Extract the boundary clipping residual sequence corresponding to the detection box and calculate its residual fluctuation within a preset time window;

[0113] For example, the volatility of the boundary clipping residual sequence is normalized statistically to reflect the stability of the occlusion boundary evolution:

[0114] Let the residual sequence corresponding to the dominant boundary be... Its mean is , Given the frame number, the residual fluctuation can be expressed as:

[0115] ;

[0116] In the formula, This refers to the residual fluctuation. The preset time window length, This is a reference scale along the direction perpendicular to the tangential direction of the occlusion. The current frame number; when the dominant boundary is the left or right boundary, The width of the predicted trajectory box can be selected; when the dominant boundary is the upper or lower boundary, The height of the predicted trajectory box can be taken; if If the value is large, it indicates that the boundary trimming evolution is unstable, and the reliability of the current observation should be further reduced;

[0117] Based on the area missing ratio, residual fluctuation and preset basic variance parameters, the positional variance value along the direction perpendicular to the occlusion tangent is generated as the observation uncertainty.

[0118] For example, the area missing percentage and residual fluctuation are used as gain terms on the basic variance parameter: Let the basic variance parameter be... The position variance along the direction perpendicular to the tangential direction of the occlusion can be expressed as:

[0119] ;

[0120] In the formula, The observation uncertainty is expressed as the positional variance along the direction perpendicular to the tangential direction of the shading, i.e., the direction of the main influence of the shading. The preset basic variance parameter is set to reflect the basic positioning error level of the system when there is no obstruction and the detection is normal. Its value range can be from 1 to 16 (unit is square pixels), and a typical value can be 4.

[0121] Furthermore, if the basic variance parameter is set too small, even if there is significant occlusion and pruning in the observations, it will lead to overly strong updates; if the basic variance parameter is set too large, it will weaken the effect of normal observations on state correction; when the proportion of missing area increases or the residual fluctuation increases, the observation uncertainty will increase synchronously, thereby automatically reducing the observation weight along the main influence direction of occlusion during state updates.

[0122] Through the above technical solution, this embodiment uses the ratio of the missing area of ​​the detection box, the residual fluctuation of the boundary clipping residual sequence, and the basic variance parameter to generate the position variance value along the direction perpendicular to the occlusion tangent. This allows the observation uncertainty to adapt to the degree of occlusion and the stability of boundary evolution, thereby suppressing error correction in the occluded direction and retaining effective observation information in other directions during the state update stage.

[0123] In an optional embodiment, after obtaining the updated trajectory state of each target in the current frame and generating the predicted trajectory box for the next frame based on it, the method further includes determining the detection box for the next frame, specifically including:

[0124] From the updated trajectory state of the current frame, determine the target that forms an occlusion phase fingerprint and a corresponding local sequential signature, and extract the next frame predicted trajectory box, continuation direction and occlusion tangent of the determined target;

[0125] Based on the predicted trajectory box of the next frame, the local interest expansion region is generated by asymmetric expansion along the continuity direction at a first preset expansion scale and along the occlusion tangent at a second preset expansion scale.

[0126] It should be noted that the predicted trajectory box of the next frame is used as the base box of the local interest expansion region, and local coordinate axes are established according to the continuation direction and the occlusion tangent. The expansion along the continuation direction first reflects the reproduction uncertainty of the target in the main occlusion influence direction, while the expansion along the occlusion tangent is used to accommodate reasonable lateral offset under the condition of local order preservation.

[0127] For example, both the first preset expansion scale and the second preset expansion scale can be defined as a proportion relative to the side length of the corresponding direction of the predicted trajectory box. The first preset expansion scale is set based on the fact that when directional occlusion is lifted or continues to develop, the target is more likely to have a large positional regression deviation along the continuing direction. The preferred value range of the first preset expansion scale is 0.15 to 0.60, and a typical value can be 0.35. The second preset expansion scale is set based on the fact that, under the premise that the local sequential signature has already provided tangential adjacency constraints, only a moderate search margin needs to be retained along the occlusion tangential direction. The preferred value range of the second preset expansion scale is 0.05 to 0.30, and a typical value can be 0.15.

[0128] To further explain, the first preset expansion scale is larger than the second preset expansion scale, so as to ensure that the local interest expansion region has a stronger fault tolerance in the continuation direction while maintaining relative convergence in the occlusion tangential direction;

[0129] In the next frame, each detection box is extracted within the local interest dilation region according to a preset extraction rule, and the center coordinates of each detection box are projected onto the occlusion tangent to determine the lateral distribution and adjacency relationship of each detection box relative to the adjacent trajectory identifiers in the local sequential signature.

[0130] It should be noted that the preset extraction rules can be used to limit the sources of detection boxes participating in the association in the next frame, and avoid introducing detection boxes that are significantly deviated from the local region of interest into the subsequent sequential screening.

[0131] For example, the preset extraction rule can be set as follows: the center of the detection box is located within the local interest expansion region, and the area overlap ratio between the detection box and the local interest expansion region meets a predetermined requirement; wherein, the overlap ratio is set based on balancing the recall rate and false detection suppression capability in the initial stage of occlusion removal, and its value range is preferably 0.30 to 0.70, with a typical value of 0.50; then, the center coordinates of the extracted detection box are projected onto the occlusion tangent, and combined with the adjacent trajectory identifier recorded in the local sequential signature, it is determined whether each detection box is located on the positive or negative side of the adjacent trajectory, and whether it maintains the nearest neighbor relationship; the purpose is to first perform region constraints, and then perform sequential relationship verification;

[0132] Detection boxes whose lateral distribution is consistent with the lateral distribution in the local sequential signature and whose predicted trajectory boxes are the closest neighbors to the adjacent trajectory labels are retained as detection boxes for the next frame.

[0133] It should be noted that a candidate detection box is only confirmed as a detection box for the target in the next frame if its lateral distribution in the occlusion tangent direction is still consistent with the local sequential signature and it maintains the nearest neighbor relationship relative to the recorded adjacent trajectories. Its purpose is that even if multiple detection boxes with close spatial positions appear in the next frame, the local sequential signature established in the previous frame and the directional search results of the current local interest expansion region can be used to jointly exclude detection candidates that do not conform to local topological continuity.

[0134] Through the above technical solution, this embodiment takes the target that forms an occlusion phase fingerprint and a corresponding local sequential signature as the object, generates a local interest expansion region around its next frame predicted trajectory box along the continuation direction and the occlusion tangent, and filters the next frame detection box in this region by combining preset extraction rules and local sequential signature, so that the detection candidates of the next frame can not only cover the reasonable reproduction range of the target in the occlusion direction, but also maintain the local sequential continuity with adjacent trajectories, thereby improving the detection hit rate and association stability during the occlusion continuation or undoing phase.

[0135] See Figure 2 As shown, this solution proposes a real-time multi-target tracking and trajectory prediction system for occluded environments, used to implement the aforementioned real-time multi-target tracking and trajectory prediction method for occluded environments, including:

[0136] The target trajectory acquisition module is used to acquire the detection box and historical trajectory status of the current frame, and determine the adjacent trajectories of each target accordingly.

[0137] The candidate association set generation module is used to generate the predicted trajectory boxes of each target in the current frame based on the historical trajectory status, and to form a candidate association set based on the spatial proximity relationship between the predicted trajectory boxes and the detection boxes in the current frame.

[0138] The phase fingerprint construction module is used to extract the indentation of the detection box relative to each boundary of the predicted trajectory box in each candidate pair in the candidate association set, and combine the indentation in the corresponding continuous frames of the historical trajectory state to determine the dominant boundary and its continuation direction, and construct the boundary clipping residual sequence to determine the occlusion phase fingerprint of the corresponding target.

[0139] The sequential signature filtering module is used to construct the occlusion tangent along the dominant boundary continuation direction when the occlusion phase fingerprint indicates that the target has entered directional occlusion. It records the relative position order of the target and the adjacent trajectory in the occlusion tangent as a local sequential signature, and retains candidate pairs that are consistent with the local sequential signature in subsequent frames to obtain a sequentially feasible candidate set.

[0140] The detection box compensation and reconstruction module is used to perform reverse geometric compensation and reconstruction on the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set according to the dominant boundary continuation direction determined by the boundary clipping residual sequence, to obtain the complete pseudo observation box of the corresponding target, and to determine the observation uncertainty along the direction perpendicular to the occlusion tangent.

[0141] The trajectory state update module is used to update the historical trajectory state of the target corresponding to the sequential feasible candidate set using the complete pseudo observation frame and observation uncertainty, and to retain the predicted trajectory frame of the current frame for targets that have not entered the sequential feasible candidate set, so as to obtain the updated trajectory state of each target in the current frame, and generate the predicted trajectory frame of the next frame based on the updated trajectory state of the current frame.

[0142] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0143] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0144] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0145] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for real-time multi-target tracking and trajectory prediction under occluded environments, characterized in that, The method includes: Obtain the current frame detection box and historical trajectory status, and determine the adjacent trajectories of each target accordingly; Based on the historical trajectory status, a predicted trajectory box for each target in the current frame is generated, and a candidate association set is formed based on the spatial proximity relationship between the predicted trajectory box and the detection box in the current frame. For each candidate pair in the candidate association set, the shrinkage of the detection box relative to each boundary of the predicted trajectory box in the candidate pair is extracted, and the dominant boundary and its continuation direction are determined by combining the shrinkage in the corresponding consecutive frames of the historical trajectory state. The boundary clipping residual sequence is constructed to determine the occlusion phase fingerprint of the corresponding target. When the occlusion phase fingerprint indicates that the target has entered directional occlusion, the occlusion tangent is constructed along the dominant boundary continuation direction. The relative position order of the target and the adjacent trajectory in the occlusion tangent is recorded as a local sequential signature. In subsequent frames, candidate pairs consistent with the local sequential signature are retained to obtain a sequentially feasible candidate set. Based on the dominant boundary continuation direction determined by the boundary clipping residual sequence, the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set are reconstructed by reverse geometric compensation to obtain the complete pseudo observation box of the corresponding target, and the observation uncertainty along the direction perpendicular to the occlusion tangent is determined. The historical trajectory states of the targets corresponding to the updated sequential feasible candidate set are obtained by using the complete pseudo observation frame and observation uncertainty, and the predicted trajectory frame of the current frame is retained for targets that have not entered the sequential feasible candidate set. The updated trajectory state of each target in the current frame is obtained, and the predicted trajectory frame of the next frame is generated according to the updated trajectory state of the current frame.

2. The method according to claim 1, characterized in that, By combining the indentation amount in consecutive frames corresponding to historical trajectory states, the dominant boundary and its continuation direction are determined. A boundary clipping residual sequence is constructed to determine the occlusion phase fingerprint of the corresponding target, specifically including: Within a preset time window, extract continuous historical detection boxes and their corresponding continuous historical predicted trajectory boxes from the historical trajectory status. Extract the boundary shrinkage of the current frame detection box relative to the current frame predicted trajectory box, and extract the historical boundary shrinkage of each consecutive historical detection box relative to its corresponding consecutive historical predicted trajectory box, and construct the boundary shrinkage sequence of each boundary in chronological order. Extract monotonically increasing gradient features from the boundary shrinkage sequence of each boundary along the time dimension. Determine the boundary with the monotonically increasing gradient features exceeding the preset deformation threshold and the largest cumulative shrinkage increment as the dominant boundary. Determine the normal direction pointing into the detection box that continuously increases the shrinkage of the boundary corresponding to the dominant boundary as the continuation direction. Extract the boundary shrinkage of the dominant boundary in the current frame and each frame within the preset time window, and construct a boundary clipping residual sequence based on the difference in shrinkage between two adjacent frames in time. The dominant boundary, continuation direction, and boundary clipping residual sequence are established as the occlusion phase fingerprint of the corresponding target.

3. The method according to claim 2, characterized in that, After extracting the monotonically increasing gradient features from the boundary shrinkage sequence of each boundary along the time dimension, the process also includes a non-formation determination process for cases where the monotonically increasing gradient features of each boundary do not reach the preset deformation threshold. Specifically, this includes: Determine whether the monotonically increasing gradient features of the corresponding target at each boundary in the current frame have not reached the preset deformation threshold. If the determination is yes, record that the corresponding target is in the state of not reaching the threshold. In the state where the threshold is not reached, the historical boundary shrinkage corresponding to the earliest historical frame within the preset time window is moved out, and the boundary shrinkage of the current frame detection box relative to the current frame prediction trajectory box is moved in, so as to update the boundary shrinkage sequence of each boundary for the determination of the dominant boundary in the next frame. Based on the state of not reaching the threshold, the occlusion phase fingerprint of the corresponding target is marked as not formed, and the process of determining the dominant boundary of the current frame is terminated.

4. The method according to claim 1, characterized in that, The relative position order of the target and its adjacent trajectories along the occlusion tangent is recorded as a local sequence signature. Candidate pairs matching this local sequence signature are retained in subsequent frames to obtain a set of sequentially feasible candidates, specifically including: Extract the center coordinates of the detection box in the current frame and the center coordinates of the predicted trajectory boxes in the current frame corresponding to each adjacent trajectory. Project the extracted center coordinates onto the occlusion tangent to obtain the corresponding one-dimensional projected coordinates. Based on the one-dimensional projection coordinates corresponding to the current frame detection box, determine the lateral distribution of each adjacent trajectory on the positive or negative side of the occlusion tangential direction, and establish the local order relationship of each side according to the distance between the one-dimensional projection coordinates of each adjacent trajectory and the current frame detection box from near to far. Extract the nearest neighboring trajectory identifier to the current frame detection box from the local order relationship on each side, and the distribution of the adjacent trajectory on the forward or reverse side, as a local order signature; In subsequent frames, the center coordinates of the candidate detection boxes of each candidate pair in the candidate association set are extracted, as well as the center coordinates of the predicted trajectory boxes corresponding to the adjacent trajectory identifiers in the local sequential signature. The extracted center coordinates are then projected onto the occlusion tangent to determine the candidate lateral distribution and candidate adjacency relationship of each candidate pair. Candidates whose lateral distribution is consistent with that in the local sequential signature are retained, and whose candidate adjacency relationship indicates that the candidate detection box and the predicted trajectory box corresponding to the adjacent trajectory label maintain the nearest neighbor relationship, constitute the sequentially feasible candidate set.

5. The method according to claim 1, characterized in that, Reverse geometric compensation reconstruction is performed on the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set to obtain the complete pseudo observation box of the corresponding target, specifically including: Extract the coordinates of the opposing boundary relative to the dominant boundary in the detection box of the feasible candidate set, as well as the coordinates of the two side boundaries intersecting with the dominant boundary; Extract three consecutive historical detection boxes from the historical trajectory state that correspond to the unoccluded phase fingerprint, and calculate their historical aspect ratio. Extract the current side length of the detection frame along the length of the dominant boundary, and combine it with the historical average width-to-height ratio to determine the compensation size along the continuity direction; Keeping the coordinates of the opposing boundary and the coordinates of the two side boundaries fixed, the coordinates of the occluded boundary are back-calculated based on the coordinates of the opposing boundary and the coordinates of the two side boundaries according to the compensation size, and then combined with the coordinates of the opposing boundary and the coordinates of the two side boundaries to form a complete pseudo observation frame.

6. The method according to claim 5, characterized in that, Determine the observation uncertainty along the direction perpendicular to the tangential direction of the obstruction, specifically including: Extract the actual observed area of ​​the detection box and the reconstructed area of ​​the complete pseudo-observation box from the feasible candidate set of the order, and determine the area missing ratio accordingly; Extract the boundary clipping residual sequence corresponding to the detection box and calculate its residual fluctuation within a preset time window; Based on the area missing ratio, residual fluctuation, and preset basic variance parameters, the positional variance value along the direction perpendicular to the occlusion tangent is generated as the observation uncertainty.

7. The method according to claim 4, characterized in that, After obtaining the updated trajectory state of each target in the current frame and generating the predicted trajectory box for the next frame based on it, the process also includes determining the detection box for the next frame, specifically including: From the updated trajectory state of the current frame, determine the target that forms an occlusion phase fingerprint and a corresponding local sequential signature, and extract the next frame predicted trajectory box, continuation direction and occlusion tangent of the determined target; Based on the predicted trajectory box of the next frame, the local interest expansion region is generated by asymmetric expansion along the continuity direction at a first preset expansion scale and along the occlusion tangent at a second preset expansion scale. In the next frame, each detection box is extracted within the local interest dilation region according to a preset extraction rule, and the center coordinates of each detection box are projected onto the occlusion tangent to determine the lateral distribution and adjacency relationship of each detection box relative to the adjacent trajectory identifiers in the local sequential signature. Detection boxes whose lateral distribution is consistent with the lateral distribution in the local sequential signature and whose predicted trajectory boxes are the closest neighbors to the adjacent trajectory labels are retained as detection boxes for the next frame.

8. A multi-target real-time tracking and trajectory prediction system under occlusion conditions, characterized in that, A method for implementing real-time multi-target tracking and trajectory prediction in an occluded environment as described in any one of claims 1-7, comprising: The target trajectory acquisition module is used to acquire the detection box and historical trajectory status of the current frame, and determine the adjacent trajectories of each target accordingly. The candidate association set generation module is used to generate the predicted trajectory boxes of each target in the current frame based on the historical trajectory status, and to form a candidate association set based on the spatial proximity relationship between the predicted trajectory boxes and the detection boxes in the current frame. The phase fingerprint construction module is used to extract the indentation of the detection box relative to each boundary of the predicted trajectory box in each candidate pair in the candidate association set, and combine the indentation in the corresponding continuous frames of the historical trajectory state to determine the dominant boundary and its continuation direction, and construct the boundary clipping residual sequence to determine the occlusion phase fingerprint of the corresponding target. The sequential signature filtering module is used to construct the occlusion tangent along the dominant boundary continuation direction when the occlusion phase fingerprint indicates that the target has entered directional occlusion. It records the relative position order of the target and the adjacent trajectory in the occlusion tangent as a local sequential signature, and retains candidate pairs that are consistent with the local sequential signature in subsequent frames to obtain a sequentially feasible candidate set. The detection box compensation and reconstruction module is used to perform reverse geometric compensation and reconstruction on the center position and width and height dimensions of the detection boxes in the sequential feasible candidate set according to the dominant boundary continuation direction determined by the boundary clipping residual sequence, to obtain the complete pseudo observation box of the corresponding target, and to determine the observation uncertainty along the direction perpendicular to the occlusion tangent. The trajectory state update module is used to update the historical trajectory state of the target corresponding to the sequential feasible candidate set using the complete pseudo observation frame and observation uncertainty, and to retain the predicted trajectory frame of the current frame for targets that have not entered the sequential feasible candidate set, so as to obtain the updated trajectory state of each target in the current frame, and generate the predicted trajectory frame of the next frame based on the updated trajectory state of the current frame.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images