A multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes

Abstract

This invention discloses a multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes, specifically relating to the fields of computer vision and intelligent perception. It constructs an initial target representation by extracting local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation order features. Furthermore, it establishes a preemptive motion corridor and generates an intersection-separation fingerprint to distinguish and constrain targets during intersection. Based on this, it obtains a unique continuation sequence of targets through candidate contraction and exclusive association. For trajectory interruptions caused by occlusion, it performs bidirectional backtracking continuation by combining boundary dissipation timing and turning coherence to generate an identity-preserving trajectory. Finally, it identifies transposition precursors based on trajectory convergence trends and performs trajectory transfer correction and output delay compensation to obtain stable multi-target real-time tracking results. This invention can improve target identity preservation and trajectory continuity in complex scenes such as strong occlusion, frequent intersections, and changes in illumination.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent perception technology, specifically to a multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes. Background Technology

[0002] In the bends and meeting areas of unmanned transport tunnels in underground mines, vehicles, pedestrians, and temporary work equipment often meet at close range within a short window where dust surges, overhead lights flicker, and wet wall reflections coexist. The targets only show discontinuous edges and local displacement signs. Most existing multi-target tracking technologies use appearance similarity combined with linear continuation to complete the association. In such small-scale, strongly occluded, and abrupt turning scenarios, it is easy to cause identity swapping after meeting, incorrect continuation after occlusion, and real-time output jitter. It is especially difficult to handle the problem of identity preservation when targets cross the same occlusion boundary one after another. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-target real-time tracking optimization algorithm suitable for complex dynamic scenarios, in order to overcome the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes, comprising: Acquire continuous scene images, extract local edge undulation features, micro-displacement direction features and occlusion boundary dissipation order features of each target to form an initial target representation; Based on the initial characterization of the target, a preemptive motion corridor is established for each target, and a convergence separation fingerprint is generated during the target convergence process based on the order of entry, turning offset, and edge yielding relationship. Specifically, based on the initial target characterization, a preemptive motion corridor is established for each target, including: Based on the center position and micro-displacement direction characteristics of the target area, a continuous displacement band is formed by connecting the center positions of the current frame and the previous two frames. The forward extension distance and lateral expansion width are determined according to the micro-displacement intensity to form the initial corridor range. The two sides of the initial corridor range are expanded or contracted according to the local edge undulation characteristics to obtain the corrected corridor. Combining the occlusion boundary dissipation order characteristics, a segmented passage priority order is set at the corridor boundary corresponding to the target contact side to form the preemptive motion corridor. During target convergence, convergence separation fingerprints are generated based on entry sequence, turning offset, and edge yielding relationships. This includes: when preemptive motion corridors overlap, recording the first frame time and entry path position of each target entering the overlapping area to form an entry sequence; statistically analyzing the directional changes in consecutive frames within the convergence area based on micro-displacement direction features to determine turning offset events and their directional shift trends; identifying the edge contraction and expansion sequence changes in the target contour contact area by combining local edge undulation features and occlusion boundary dissipation order features to form edge yielding relationships; and encoding the entry sequence, turning offset events, and edge yielding relationships in chronological order to form a convergence separation fingerprint. The preemptive motion corridor is used to narrow down the candidates for targets at adjacent time points, and the intersection and separation fingerprint is used to perform exclusive association to obtain a unique continuation sequence of targets; For the interrupted segments in the unique continuation sequence, bidirectional backtracking is performed based on the occlusion boundary dissipation order characteristics and the turning coherence before and after the interruption to generate an identity-preserving trajectory. Based on the identity-preserving trajectory, the trajectory convergence trend features within consecutive frames are extracted. When a transponder precursor is detected, trajectory transfer correction and output delay compensation are performed to obtain real-time tracking results for multiple targets.

[0005] Preferably, acquiring continuous scene images includes: Brightness normalization and median filtering are performed on continuous scene images to obtain preprocessed images; foreground candidates are extracted based on the grayscale difference between the current frame and the previous 1 frame and 2 frames, and connected regions are formed after morphological processing; the connected regions are filtered according to area, aspect ratio and boundary integrity to obtain target regions.

[0006] Preferably, extracting local edge undulation features includes: Each target region is expanded outwards according to the minimum bounding rectangle to form a local analysis range; Extract the outermost closed contour within the local analysis range, and divide the closed contour into 8 equal segments according to length; For each contour segment, the maximum outward distance, the maximum absolute inward distance, the number of distance sign changes, and the average absolute distance are calculated to form local edge undulation features.

[0007] Preferably, the extraction of micro-displacement direction features includes: Comparison blocks are placed within each target area, and point-by-point matching is performed around the corresponding position in the previous frame to obtain the micro-displacement of each comparison block; comparison blocks with grayscale variance lower than the preset value or unstable matching are removed, and valid comparison blocks are retained. The micro-displacements of all valid comparison blocks are statistically analyzed by direction to determine the principal direction. The micro-displacement intensity is determined based on the displacement length, and the lateral drift is determined based on the proportion of offset perpendicular to the principal direction, thus forming the micro-displacement direction characteristics.

[0008] Preferably, the extraction of occlusion boundary dissipation order features and the formation of an initial target representation include: For neighboring target pairs, the contact side is determined and the contact side contour is divided into 4 continuous boundary segments. The boundary point density, inner and outer grayscale contrast and continuous length ratio of each boundary segment are calculated in consecutive frames, and the boundary visibility value is obtained accordingly. The dissipation order is determined according to the order in which the boundary visibility value of each boundary segment decreases to the preset proportion of the average boundary visibility value before occlusion, and the occlusion boundary dissipation order feature is formed by combining the recovery order. The local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation order features are written in a fixed field order to form the initial characterization of the target.

[0009] Preferably, the target unique continuation sequence is obtained, including: The preemptive motion corridor of the previous frame is used to perform spatial projection screening of the target in the current frame, and the direction consistency screening is combined with the micro displacement direction features to form a candidate set. The intersection and separation fingerprints in the candidate set are compared with the intersection and separation fingerprints in the historically confirmed continuation sequence one by one, and the candidate targets with conflicts are eliminated. The remaining candidate targets are sorted according to the distance from their center position to the central axis of the preemptive motion corridor, and the continuing targets are determined by combining the inter-frame interval information and micro-displacement intensity. The continuity of the target is confirmed by continuous frame consistency verification to obtain a unique continuity sequence of the target.

[0010] Preferably, generating an identity-preserving trajectory includes: The interrupted segments in the unique continuous sequence are identified, and the center position of the target area, the micro-displacement direction features, and the dissipation order features of the occlusion boundary of the effective frames before and after the interruption are recorded. The occlusion boundary dissipation order features of candidate targets within the interrupted segment are compared along the reverse and forward time directions to determine segments with consistent dissipation order. Based on the micro-displacement direction characteristics before and after the interruption, the turning continuity section is screened, and the continuation point is selected within the intersection range of the dissipation sorting consistent section and the turning continuity section. Insert the reconnection point into the target's unique continuation sequence and perform continuous frame consistency verification to generate an identity-preserving trajectory.

[0011] Preferably, when a premonition of a switching event is detected, trajectory handover correction and output delay compensation are performed to obtain real-time multi-target tracking results, including: Based on the changes in the distance between the center positions of the target region and the target area within consecutive frames, the trajectory convergence trend features are extracted. Select all targets with trajectory convergence trends, calculate the spatial overlap between two targets in the corresponding frame, and determine that the two targets are in a spatially close state when their trajectory convergence regions have overlapping pixels in the same frame and the overlapping area is greater than 15% of the area of ​​either target region. Compare the micro-displacement direction features of two targets, convert the difference in direction number into a difference in direction level, and determine that the direction is consistent when the difference does not exceed one direction level in two consecutive frames. Only when both the spatial proximity state and the orientation consistency state are met simultaneously is it determined that there is a premonition of crosstalk between the target pairs, and the corresponding frame number and the number of the participating target are recorded. Based on the historical sequence of entry and the marginal yielding relationship, the dominant and subordinate objectives are determined, and the execution trajectory of the subordinate objectives is transferred and corrected. For subordinate targets that undergo trajectory transfer correction, the average position of the center position of consecutive frames before correction is used for compensation output within a preset number of frames to obtain multi-target real-time tracking results.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a joint constraint structure of "preemptive motion corridor + convergence-separation fingerprint" to uniformly encode the continuous motion trajectory, boundary contact process, and occlusion dissipation sequence of a target in physical space. This transforms target association from relying solely on single-frame appearance or positional similarity into a matching problem of motion channel occupancy relationships and boundary evolution timing. The preemptive motion corridor essentially defines the target's reachable spatial range within a short timeframe, constrained by historical displacement paths and local contour deformations, reflecting the target's inertial motion and spatial exclusivity in a real physical scene. The convergence-separation fingerprint, by recording the target's entry sequence, turning changes, and boundary yielding behavior during convergence, characterizes the dynamic interaction relationships between targets. Compared to existing methods based solely on distance or appearance matching, this invention, starting from the physical contact and motion evolution mechanism, effectively suppresses the identity swapping problem in convergence scenarios, fundamentally improving the determinism and stability of multi-target association.

[0013] 2. This invention introduces a temporal recovery mechanism of "occlusion boundary dissipation order + bidirectional backtracking continuation + trajectory convergence trend correction," using the changes in boundary visibility and motion continuity during the occlusion process as core criteria to achieve self-consistent reconstruction of the target trajectory in the time dimension. Specifically, the occlusion boundary dissipation order reflects the sequential changes in each boundary when the target is occluded and then re-revealed, essentially stemming from the target's relative occlusion relationship in space. Bidirectional backtracking continuation utilizes the consistency of motion direction and position convergence characteristics before and after occlusion to perform a squeezing-style matching of the interrupted trajectory, thereby restoring the true motion path. The trajectory convergence trend is further used to identify target approach and potential switching states in advance, and positional abrupt changes are eliminated through trajectory transfer and time delay compensation. Unlike existing processing methods that rely on simple interpolation or short-term prediction, this invention, based on the constraints of the occlusion physical process and motion continuity, effectively repairs trajectory breakage and jump problems in complex occlusion scenarios, directly improving the continuity and spatial consistency of tracking results. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a multi-target real-time tracking optimization algorithm applicable to complex dynamic scenarios according to the present invention.

[0016] Figure 2 This is a flowchart of the fingerprint generation method for intersection separation according to the present invention.

[0017] Figure 3 This is a flowchart of the identity preservation trajectory generation method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] For examples, please refer to Figure 1 As shown in the figure, the multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes described in this embodiment includes: Acquire continuous scene images, extract local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation order features of each target to form an initial target representation.

[0020] In this embodiment, the continuous scene images are obtained from a fixedly installed video acquisition device, with a sampling frequency of 30 frames per second and a single frame resolution of 1920 x 1080. During acquisition, the current frame, the previous frame, the previous two frames, the previous three frames, and the previous four frames are saved in chronological order to form a continuous image sequence of 5 frames in length, so as to extract the chronological relationship of minute displacements and boundary dissipation processes later.

[0021] To minimize the impact of illumination fluctuations on subsequent processing, each frame undergoes brightness normalization. Specifically, the entire frame is divided into 16 equal-sized regions, and the average grayscale value of each region is calculated. The average grayscale value of each region is then adjusted to a range close to the average grayscale value of the entire frame. If the difference between the average grayscale value of a region and the average grayscale value of the entire frame exceeds 20 grayscale levels, each pixel within that region is compensated for the same difference. After brightness normalization, a 3x3 neighborhood median filter is used to remove isolated noise points, resulting in a preprocessed image for target extraction.

[0022] After obtaining the preprocessed image, the target region is first extracted from the continuous image sequence. Specifically, the pixel-by-pixel grayscale difference between the current frame and the previous frame, and between the current frame and two previous frames, is calculated. The two grayscale difference results are then averaged to obtain the inter-frame variation map. To avoid false detections caused by local flicker, pixels with less than 18 grayscale levels in the inter-frame variation map are set to zero, while pixels with 18 or more grayscale levels are retained as foreground candidates.

[0023] A 5x5 neighborhood closing operation and a 3x3 neighborhood opening operation are performed on the foreground candidates to connect broken regions and remove minor glitch. The area, aspect ratio, and boundary integrity of each processed connected region are calculated. Regions with an area less than 80 pixels, regions with an aspect ratio less than 0.2 or greater than 5, and regions with a boundary integrity lower than 0.4 are removed. Boundary integrity is calculated based on the relationship between the length of the region's circumscribed contour and the region's area; regions with excessively fragmented contours have lower boundary integrity.

[0024] The connected regions retained after filtering are used as target regions in the current frame. For cases where the distance between the center points of two adjacent target regions is less than 20 pixels and the minimum spacing between their contours is less than 6 pixels, they are recorded as nearest neighbor target pairs for use in subsequent extraction of occlusion boundary dissipation order features.

[0025] After obtaining the target region, local edge undulation features are extracted. Specifically, each target region is first expanded outward by 6 pixels according to its minimum bounding rectangle to form a local analysis range; then, a 5x5 neighborhood Gaussian smoothing is performed within this local analysis range to reduce the disturbance of texture noise on the edge position.

[0026] After smoothing, the horizontal and vertical grayscale changes of each pixel are calculated and combined to form the edge intensity. To determine valid edges, the median value of all edge intensities within the local analysis range is first calculated. Then, the high threshold is set to 1.5 times the median value, and the low threshold is set to 0.4 times the high threshold. Edge points above the high threshold are directly retained, and edge points between the low and high thresholds that are connected to the retained edge points are also retained. The remaining edge points are discarded. After obtaining the edges, the outermost closed contour is extracted and divided into 8 segments according to the contour length. The undulation information of each segment is calculated separately. A sliding window with a length of 21 contour points is used in each contour segment. A reference line is drawn connecting the first and last points of the window, and the distances from the remaining contour points in the window to this reference line are calculated. The distance towards the target centroid is recorded as the inward distance, and the distance away from the target centroid is recorded as the outward distance. For each sliding window, the maximum value of the outward distance, the maximum absolute value of the inward distance, the number of times the distance sign changes, and the average value of all absolute distances are recorded.

[0027] The above four results reflect the degree of convexity, indentation, frequency of sawtooth variation, and overall undulation intensity of the local contour, respectively. Arranging the four results of the eight contour segments in a fixed order yields the local edge undulation characteristics of the target.

[0028] After obtaining the local edge undulation features, the micro-displacement direction features are extracted. Specifically, for each target region, 16 comparison blocks are evenly distributed within it. Each comparison block is 8 by 8 pixels in size, and the center of the comparison block is at least 4 pixels away from the boundary of the target region to avoid the direct impact of boundary jitter.

[0029] For each comparison block in the current frame, the most similar position is found point-by-point within a 9x9 pixel search range around the corresponding position in the previous frame. The similarity is calculated as follows: the absolute value of the grayscale difference between the corresponding pixels of the current frame comparison block and a candidate block within the search range is taken, and then the absolute values ​​of these 64 pixels are summed. The candidate block with the smallest sum is taken as the matching position. The displacement between the center of the current frame comparison block and the center of the matching position is the micro-displacement of that comparison block.

[0030] To eliminate false matches caused by insufficient texture, the internal grayscale variance of each comparison block is recalculated. Comparison blocks with a grayscale variance of less than 15 are considered invalid comparison blocks and are not included in the statistics. At the same time, when the sum of the minimum grayscale differences of a certain comparison block is greater than 0.7 times the average value of this item of all comparison blocks, it is also considered that the match is unstable and is not included in the statistics.

[0031] The micro-displacements of all valid comparison blocks are divided into eight directions, and the number of upward, downward, leftward, rightward, and four diagonal directions is counted. The direction with the most counts is designated as the principal direction. The average displacement length of each valid comparison block is then calculated and categorized into four levels: 0-1 pixel, 1-3 pixels, 3-5 pixels, and greater than 5 pixels, representing the micro-displacement intensity. For the component perpendicular to the principal direction, the proportion of positive and negative offsets is calculated to characterize the lateral drift of the target during its movement along the principal direction. The principal direction, micro-displacement intensity, and lateral drift proportion are then combined to form the micro-displacement direction feature of the target.

[0032] After obtaining the micro-displacement direction features, the occlusion boundary dissipation order features are extracted. This feature is calculated only for the aforementioned nearest neighbor target pairs.

[0033] In the current frame, for two neighboring target regions, first find the point pair with the smallest distance between their two outer contours, and determine the contact side based on the location of this point pair. Then, using the main direction of the target as a reference, divide the contact side contour from the end facing the movement to the end facing away from the movement into four consecutive boundary segments, each boundary segment covering one-quarter of the total length of the contact side contour. For each boundary segment, calculate the boundary visibility value over five consecutive frames. The boundary visibility value consists of three parts: the first part is the boundary point density, which is the ratio of the number of effective edge points in the boundary segment to the length of the boundary segment; the second part is the inner and outer grayscale contrast, which is calculated by taking three wide-band regions on each side of the boundary segment, calculating the average grayscale, and then taking the absolute value of the difference between the two; the third part is the continuous length ratio, which is the ratio of the longest continuous edge length in the boundary segment to the total length of the boundary segment. To make the three results comparable, the boundary point density, inner and outer grayscale contrast, and continuous length ratio were first converted to integer values ​​from 0 to 100. The average value of the segment in the first three frames before occlusion was used as the baseline of 100. Values ​​below half the baseline were recorded as 50, values ​​below one-quarter of the baseline were recorded as 25, and so on.

[0034] After conversion, the boundary visibility value is obtained by allocating 50 parts to boundary point density, 30 parts to internal and external grayscale contrast, and 20 parts to continuous length ratio.

[0035] If the boundary visibility value of a certain boundary segment is lower than 60% of the average boundary visibility value of the previous three unobstructed frames for two consecutive frames, then the frame in which the boundary segment first falls below this standard is recorded as the dissipation time. The four boundary segments are sorted from earliest to latest according to their dissipation time to obtain the dissipation order of the occluded boundaries; if there are cases where the dissipation time is the same, the boundary segment with the larger decrease in boundary visibility value is sorted first.

[0036] For the re-reveal process after the occlusion ends, the same method is used to record the moment when each boundary segment recovers to 70% of the average boundary visibility value of the three frames before the occlusion ends, and the order of recovery is saved as additional content of the occlusion boundary dissipation sequence feature.

[0037] After obtaining the local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation sequence features, an initial target characterization is formed.

[0038] First, the distance values ​​in the local edge undulation features are scaled and normalized. The normalization is based on the length of the short side of the smallest bounding rectangle of the target. The maximum outward distance, the maximum absolute inward distance, and the average of all absolute distance values ​​are divided by the length of the short side and then converted to integers from 0 to 255. For the number of distance sign changes, the frequency of each 20 contour points is converted to integers from 0 to 255.

[0039] The main direction in the micro-displacement directional characteristics is represented by direction numbers from 1 to 8, the micro-displacement intensity is represented by levels from 1 to 4, and the lateral drift ratio is represented by integers from 0 to 100.

[0040] The four boundary segments in the dissipation order feature of the occlusion boundary are numbered 1, 2, 3, and 4 according to the end facing the motion to the end facing away from the motion. Their dissipation order is directly written as a 4-bit sequence code, and the recovery order is written as another set of 4-bit sequence codes. If the current target does not form a neighboring target pair with other targets, this item is written as 0, 0, 0, 0, and marked as "no occlusion" in the status bit.

[0041] The target number, timestamp, target region center position, target region width and height, local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation order features are continuously written in a fixed field order to form the target initial representation corresponding to the target in the current frame.

[0042] The initial representation of the target is used in subsequent steps to perform preemptive motion corridor establishment, convergence separation fingerprint construction, and target identity continuation judgment. The field order of the same target remains consistent in consecutive frames to ensure that feature comparisons at different times have a unified basis.

[0043] In practice, to ensure the stability of the above feature extraction results, the initial target representation obtained from three consecutive frames of the same target can be subjected to a time smoothing process.

[0044] For the integer values ​​in the local edge undulation features, the average is calculated by taking 50 parts from the current frame, 30 parts from the previous frame, and 20 parts from the previous two frames, and then rounding down. For the main direction in the micro-displacement direction features, it is determined by a majority vote of 3 frames. If the directions are different in all 3 frames, the direction of the current frame is retained. For the micro-displacement intensity and the proportion of lateral drift, the same weighted average method as the local edge undulation features is used. For the occlusion boundary dissipation order features, they are only written into the initial representation of the formal target when the sorting results of 2 consecutive frames are consistent. Otherwise, they are temporarily stored in a pending confirmation state.

[0045] Through the above processing, the resistance to jumps caused by single-frame noise can be improved without changing the definition of each feature, so that the initial target representation can stably reflect the local contour changes, short-term displacement direction, and the boundary dissipation sequence during the occlusion contact process.

[0046] Please see Figure 2 As shown, based on the initial target characterization, a preemptive motion corridor is established for each target, and during the target convergence process, convergence separation fingerprints are generated according to the order of entry, turning offset, and edge yielding relationship.

[0047] First, based on the micro-displacement direction characteristics in the initial characterization of the target and the center position of the target region, a continuous displacement zone is constructed.

[0048] Read the center position of the target region in the current frame and the center positions of the corresponding targets in the previous two frames. Connect the center positions of the three frames in chronological order to form a polyline trajectory. Use this polyline trajectory as the main direction reference line and perform point-by-point interpolation on the path between the center position of the current frame and the center position of the previous two frames. Set the interpolation interval to 1 pixel distance to obtain a continuous path point sequence.

[0049] Based on the consistency of the main direction in the micro-displacement direction characteristics, if the directions of the three frames are consistent, the direction of the current frame is used as the extension direction; if there is a deviation, the direction that appears most frequently is taken as the main direction. The forward extension distance is determined according to the micro-displacement intensity: when the micro-displacement intensity level is 1, the forward extension distance is set to 1 times the length of the short side of the current target area; when the level is 2, it is set to 1.5 times; when the level is 3, it is set to 2 times; and when the level is 4, it is set to 2.5 times. The lateral expansion width is determined by the target area width and the micro-displacement intensity, specifically by multiplying the target area width by the corresponding intensity coefficient, with coefficients of 0.6, 0.8, 1.0, and 1.2 respectively.

[0050] Using the path point sequence as the central axis, the lateral width is symmetrically expanded at each path point along a direction perpendicular to the main direction. All expanded areas are merged to form a continuous strip-shaped area, i.e., the initial corridor range.

[0051] The initial corridor range is adjusted based on local edge undulation features. Specifically, the edges obtained from the target contour are spatially mapped to the initial corridor boundary, and the projection area of ​​each edge segment on the initial corridor boundary is marked.

[0052] For edge segments with large outward distances, first calculate the ratio of the maximum outward distance to the length of the target's short side. When this ratio is greater than 0.15, extend the corridor boundary of the corresponding projection area outward, with the extension distance being 80% of the maximum outward distance as the pixel value. When the ratio is less than or equal to 0.15, the extension distance is 50% of the maximum outward distance. For edge segments with large inward distances, calculate the ratio in the same way, and shrink the corresponding area boundary inward according to the above proportions.

[0053] For edge segments with a high number of distance symbol changes, when the number of changes exceeds 5 times out of every 20 contour points, the corresponding boundary is not expanded or contracted, but only its original position is maintained to avoid the unstable boundary area from having an excessive impact on the corridor shape. After processing all edge segments, the adjusted corridor boundary is smoothed. Specifically, a set of points is taken every 10 pixels along the boundary line, and the average of the boundary positions of three adjacent sets of points is taken as the new boundary point, thus obtaining a corrected corridor that fits the changes in the target contour.

[0054] Based on the dissipation order characteristics of the occlusion boundary, a segmented passage priority order is set on the target contact side.

[0055] Determine the corridor boundary region where the target contact side is located, and divide the contact side boundary into 4 continuous sub-regions according to the 4 boundary segments given in the occlusion boundary dissipation sequence feature. The length of each sub-region is consistent with the corresponding boundary segment.

[0056] For sub-regions with higher dissipation order, they are marked as priority passage areas, and their priority values ​​are recorded. These values ​​are assigned sequentially from 1 to 4, with 1 representing the highest priority. For non-priority areas, a delay restriction is implemented: when a target enters this area, the entry frame number is recorded, and the target must remain in this area for at least two consecutive frames before continuing forward. If this condition is not met, the target's position update within the corridor is frozen, maintaining the position from the previous frame. For priority passage areas, no such restriction is applied, allowing the target to move continuously in micro-displacement directions. This method creates a spatially ordered structure within the corridor with a sequential occupancy relationship.

[0057] Perform consecutive frame consistency checks on the preemptive motion corridor. Take the preemptive motion corridor regions of the current frame and the previous frame, and the current frame and the two previous frames respectively, and calculate the overlap ratio between each pair.

[0058] The overlap ratio is calculated as follows: count the number of pixels in the overlapping part of the two corridor areas, then count the total number of pixels in each of the two corridor areas, take the smaller total number of pixels as the benchmark, divide the number of overlapping pixels by the benchmark, and convert the result into a percentage value.

[0059] When the overlap ratio of any group is less than 60%, the consistency between the current corridor and the historical corridor is deemed insufficient. In this case, the current corridor result is discarded, and the continuous displacement band construction, boundary adjustment, and priority setting steps are re-executed based on the initial target representation of the current frame. When the overlap ratio of both groups is not less than 60%, the current corridor is retained and used as the reference corridor for the next frame. Through this verification process, the preemptive motion corridor maintains continuity and stability in the time dimension.

[0060] The order of entry is determined based on the preemptive motion corridor. Overlapping regions are detected in the preemptive motion corridors of consecutive frames. When the corridor regions of two or more targets overlap pixels in the same frame and the overlap area is greater than 10% of their respective corridor areas, it is determined that they have entered the intersection process.

[0061] For each target entering the intersection, the frame number of its first appearance in the corridor overlap is recorded as the first frame time, and the position of the target's center point in the corridor within that frame is extracted. The corridor width is divided into three equal parts horizontally, with the middle third defined as the central axis region and the two sides as edge regions. If a target's center point is located in the central axis region within the first frame time, and its first frame time is at least one frame earlier than other targets, it is marked as the preemptive target; the remaining targets are marked as following targets. If there are targets with the same first frame time, the entry path length is compared. The path length is obtained by accumulating the total displacement distance of the center point from the previous two frames to the current frame, and the target with the longer path is marked as the preemptive target.

[0062] Based on the above determination results, an entry sequence is formed in chronological order, and the entry position category and corresponding frame number of each target are recorded.

[0063] The steering offset event is calculated based on the micro-displacement direction characteristics in the initial target representation. Within the intersection zone, at least three frames of micro-displacement direction are continuously recorded for each target. The direction number of each frame is converted into an angle interval, with eight directions corresponding to eight equally divided intervals within the range of 0 degrees to 360 degrees. The change in direction between two adjacent frames is calculated, specifically the angle difference corresponding to the difference in direction numbers between the two frames, and the smallest rotation direction is taken as the change. When the change in direction for two consecutive frames is greater than 45 degrees, a steering offset event is determined to have occurred, and the starting frame number of the event is recorded.

[0064] Regarding the directional offset trend, if the directional number increases or decreases continuously, it is recorded as a unidirectional offset; if the directional number increases first and then decreases or decreases first and then increases, it is recorded as a swing offset.

[0065] To avoid misjudgments caused by occasional jitter, if the directional change meets the condition of being greater than 45 degrees in only 2 out of 3 consecutive frames, it is necessary to further determine whether the displacement lengths corresponding to these 2 frames are both greater than 0.6 times the average displacement length of all valid comparison blocks. If so, it is still determined to be a steering offset event. All steering offset events of each target within the intersection area are recorded in chronological order, with the corresponding directional offset trend appended.

[0066] Edge yielding relationships are identified by combining local edge undulation features with occlusion boundary dissipation order features. Specifically, within the frame range determined to be intersecting, the contact area between two target contours with a minimum distance of less than 5 pixels is extracted, and edge changes are analyzed within this area.

[0067] For each target, in the boundary segment corresponding to the contact area, the maximum outward distance and the maximum absolute inward distance are extracted from the local edge undulation features and compared over three consecutive frames. If the maximum absolute inward distance of a target in the contact area increases by more than 30% of its average value in the unoccluded state over two consecutive frames, and the corresponding boundary segment is ranked low in the occlusion boundary dissipation order feature, then the target is determined to exhibit boundary contraction behavior. If another target's maximum outward distance increases by more than 20% of its average value in the unoccluded state over the same time period, then it is determined to exhibit boundary expansion behavior.

[0068] The target exhibiting boundary contraction behavior first and dissipating later in the sequence is marked as the one actively yielding, and the other target is marked as the one passing through. If both targets simultaneously contract or expand, their magnitudes of change are compared, and the one with the larger magnitude is used as the criterion. The above determination results are recorded in chronological order throughout the entire convergence process, forming a boundary yielding relationship sequence.

[0069] The sequence of entry, the turning bias event, and the edge yielding relationship are uniformly encoded to form the intersection and separation fingerprint.

[0070] Set a sequence structure with a fixed length of 12 fields. The first 4 fields are used to record the order of entry, and each field contains the target number, entry frame number and position category. The middle 4 fields are used to record the turning offset event, and each field contains the event start frame number, duration frame number and direction offset trend. The last 4 fields are used to record the edge yielding relationship, and each field contains the judgment frame number and yielding role mark.

[0071] If the actual number of records is less than the number of corresponding fields, it is padded with 0; if it exceeds the number of fields, the earliest record is retained in chronological order. For inter-frame interval information, it is represented by the difference between the frame numbers of two adjacent records, and this difference is appended to the corresponding field. All fields are arranged in chronological order to form a complete set of intersection and separation fingerprints, which are then associated with the corresponding target number and stored for exclusive determination in the subsequent target association process.

[0072] By following the steps above, target relationships during the intersection process can be distinguished and identified without relying on appearance similarity.

[0073] The candidate shrinking of targets at adjacent time moments is performed using the preemptive motion corridor, and the exclusive association is performed using the intersection separation fingerprint to obtain the unique continuation sequence of the targets.

[0074] First, candidate shrinkage processing is performed based on the preemptive motion corridor. For each target in the current frame, the preemptive motion corridor range corresponding to it in the previous frame is read. The center position of the target region in the current frame is mapped to the corridor coordinate system of the previous frame to determine whether it falls within the corridor range. If the target center position is located inside the corridor boundary, it is considered as an initial candidate; if it is located outside the boundary but less than 5 pixels away from the boundary, the shortest distance from it to the corridor centerline is calculated. If this distance is less than 0.8 times the lateral unfolding width of the corridor, it is still retained as a candidate; other cases are discarded.

[0075] Simultaneously, the retained candidate targets undergo further directional consistency screening. The deviation between the micro-displacement directional features of the target in the current frame and the main corridor direction is compared. Eight directions are categorized into adjacent relationships. If the deviation exceeds two directional levels, it is determined to be directionally inconsistent and removed from the candidate set. After this dual spatial and directional screening, the candidate set is obtained.

[0076] To avoid the impact of short-term jitter, if a target meets the candidate conditions in the first two frames but not in the current frame, it is allowed to retain its status as a continuation candidate for one frame and is marked as pending confirmation.

[0077] Exclusive association processing is performed based on the intersection and separation fingerprint. For each target in the candidate set, its corresponding intersection and separation fingerprint is extracted and compared item by item with the fingerprints in the historically confirmed continuation sequence. The comparison includes the entry sequence, turning offset events, and edge yielding relationships. For the entry sequence, the entry order of the candidate targets in the current time interval is compared with the historical records. If the order is reversed within the same time period, it is determined to be a conflict. For turning offset events, the candidate targets are compared to see if there are any unrecorded offset events in the corresponding frame range, or if the directional offset trends of existing events are consistent. If the difference exceeds one directional level, it is determined to be a conflict. For edge yielding relationships, the role markings of the candidate targets in the contact area are compared with the history. If the active yielding party and the passing party are swapped, it is determined to be a conflict.

[0078] If any conflict occurs, the candidate objective is removed from the candidate set. For candidate objectives without conflict, they are retained as part of the exclusive shrinking candidate set.

[0079] To improve the stability of the judgment, when a candidate target has only a slight deviation in one item, it is necessary to further compare its inter-frame interval information. If the interval difference is less than 2 frames, it is allowed to be retained; otherwise, it is still rejected.

[0080] Perform optimal association on the candidate set after exclusive shrinking. For each target to be associated, sort the targets in the candidate set according to the distance from their center position to the central axis of the preemptive motion corridor.

[0081] Draw a perpendicular line from the center point of the target to the central axis of the corridor, and record the length of this perpendicular line as the distance value. Sort all candidate targets in ascending order of distance value, and select the one with the smallest distance as the priority candidate.

[0082] If there are targets with the same distance, further compare the inter-frame interval information and prioritize targets with the same time interval as those in the historical continuous sequence; if they still cannot be distinguished, compare their micro-displacement intensity and select the target with the smallest intensity difference from the previous frame.

[0083] After identifying the priority candidate, it is associated with the target in the previous frame, and the associated frame number and corresponding position changes are recorded. The remaining candidate targets are not associated for the time being and are reserved for the next round of judgment.

[0084] The established association relationships are validated for consistency across consecutive frames to form a unique continuation sequence of targets. The association results of the same target within 3 consecutive frames are checked. If the same target number is associated with the same target number in all 3 frames, the association relationship is confirmed to be valid and written into the unique continuation sequence of targets. If an association is interrupted or the target number changes once within 3 frames, the association segment is marked as unstable and the process is backtracked to the candidate shrinking step to re-execute the filtering and association processing.

[0085] During the backtracking process, only the frames that have changed are recalculated, while the remaining stable frames remain unchanged. Through this consistency verification, short-term mismatches can be filtered out, ensuring that the final target unique continuation sequence remains continuous in time and consistent in identity, providing a reliable foundation for subsequent trajectory construction.

[0086] Please see Figure 3 As shown, for the interrupted segment in the unique continuation sequence, bidirectional backtracking is performed based on the occlusion boundary dissipation order characteristics and the turning continuity before and after the interruption to generate an identity-preserving trajectory.

[0087] The continuity of target numbers is checked frame by frame in the unique continuous sequence. When no correlation is formed between two adjacent frames, the position is marked as the start of the interruption, and the search continues until a frame with a correlated target reappears as the end of the interruption, thus determining a complete interruption segment.

[0088] For each interrupted segment, the center position of the target region, the micro-displacement direction features, and the occlusion boundary dissipation sequence features of the frame preceding the interruption start point are recorded, along with the corresponding information of the frame following the interruption end point. The time range of the interrupted segment is defined as all frames from the interruption start frame number plus 1 to the interruption end frame number minus 1.

[0089] To avoid mismatches caused by excessively long time intervals, when the length of an interrupted segment exceeds 5 frames, the segment is divided into multiple sub-segments with a length not exceeding 5 frames, and subsequent processing is performed on each sub-segment.

[0090] Bidirectional backtracking matching is performed based on the occlusion boundary dissipation order features. Starting from the previous valid frame, the search proceeds backward in time, frame by frame, into the interrupted segment. Simultaneously, starting from the subsequent valid frames, the search proceeds forward in time, frame by frame, into the interrupted segment.

[0091] In each search frame, the occlusion boundary dissipation order features of candidate targets are extracted and compared with the features recorded before and after the interruption. The comparison method is as follows: the dissipation order of the four boundary segments is aligned position by position, and the number of identical positions is counted. When the number of identical positions is not less than 3, the order is considered consistent. For each candidate target that meets the order consistency condition in each frame, its frame number and corresponding position are recorded, and consecutive frames that meet the condition are merged to form a dissipation order consistency segment.

[0092] If the segments obtained from the reverse search and the forward search overlap, the overlapping segment is selected as the candidate region; if there is no overlap, the segment with the longer length is selected as the candidate region. This process limits the time range within which boundary changes are consistent during the occlusion process.

[0093] Candidate regions are filtered based on the continuity of steering before and after the interruption. Micro-displacement direction features are extracted from the frame before the interruption start point and the frame after the interruption end point, and their direction numbers are converted into angle interval representations. For each frame in the candidate region, the difference between its direction number and the directions before and after the interruption is calculated. The difference is calculated as the minimum rotation amount corresponding to the angle difference between the two direction numbers. When the candidate frame direction simultaneously satisfies that the difference with the direction before the interruption does not exceed two direction levels, and the difference with the direction after the interruption does not exceed two direction levels, the frame is marked as a directionally continuous frame.

[0094] Simultaneously, it is determined whether the directional change trend is consistent. If the candidate frame's direction shows a unidirectional increasing or decreasing change in two consecutive frames, and is consistent with the directional change trend before and after the interruption, then the frame is retained; otherwise, it is discarded. All frames that meet the conditions are combined into a direction-coherent segment, and the intersection of this segment with the previously mentioned dissipation sorting segment is taken to obtain the final candidate segment.

[0095] Then, reconnection point selection is performed within the final candidate segment. For each frame in the candidate segment, the distance between the center position of its target region and the center position of the frame before the interruption, as well as the distance between the center position of the frame after the interruption, are calculated.

[0096] The distance is calculated by statistically analyzing the pixel differences between two center points in the horizontal and vertical directions, squaring both values, summing them, and taking the square root of the sum. For each candidate frame, the distance trend between it and the previous and next frames is calculated. When the distance gradually decreases or approaches stability over time, it is considered that the position has converged.

[0097] Frames that simultaneously satisfy the conditions of directional continuity, consistent dissipation order, and positional convergence are preferentially selected as continuation points. If multiple frames meet the conditions, the frame with the smallest average position before and after the interruption is selected as the final continuation point. If multiple consecutive frames in an interrupted segment meet the conditions, all of them are selected in chronological order to form a continuation path.

[0098] Insert the continuation point into the unique continuation sequence and confirm consistency.

[0099] The selected continuation points are inserted frame by frame into the interruption position corresponding to the original unique continuation sequence in chronological order, and each inserted frame is assigned the same target number as the target before the interruption.

[0100] After insertion, a consistency check is performed on three consecutive frames containing the continuation point. Specifically, the check is performed to see if the target number is consistent within the three frames and if the change in its micro-displacement direction is continuous. If the conditions are met, the continuation is confirmed to be valid. If there are inconsistent numbers or changes in direction exceeding two directional levels, the continuation is determined to have failed, and the continuation point is removed. The candidate segment is then returned to re-execute the filtering and selection steps.

[0101] Through the above processing, the interrupted segments are reasonably filled in time and their identities are kept consistent with the original target, thus forming a complete identity preservation trajectory.

[0102] Based on the identity-preserving trajectory, the trajectory convergence trend features within consecutive frames are extracted. When a transponder precursor is detected, trajectory transfer correction and output delay compensation are performed to obtain real-time tracking results for multiple targets.

[0103] The center positions of the same target in consecutive frames are recorded, with at least three consecutive frames selected as the analysis interval. The center position distances between the first and second frames, and between the second and third frames, are calculated separately. The distance is calculated as follows: the pixel differences in the horizontal and vertical directions between the two center points are counted, the squares of both are summed, and the square root of the result is taken.

[0104] Comparing the two distances mentioned above, a convergence change is recorded when the latter distance is less than the former distance, and the difference between the two distances is greater than 20% of the average distance of the first three frames. If at least two convergence changes occur in three consecutive frames, the target is determined to have a trajectory convergence trend.

[0105] Simultaneously, the starting frame number of the convergence trend and the corresponding center position change path are recorded. To ensure the stability of the judgment, the average distance of the first 3 frames is obtained by summing the distances between each pair of the 3 consecutive frames and then dividing by 2, and the integer value is taken as the benchmark.

[0106] Within the same time interval, all targets exhibiting trajectory convergence trends are selected and compared pairwise. First, the spatial overlap between two targets within the corresponding frame is calculated. When the trajectory convergence regions of two targets have overlapping pixels within the same frame and the overlapping area is greater than 15% of the area of ​​either target region, they are considered to be in a spatially close state.

[0107] Subsequently, the micro-displacement direction characteristics of the two targets are compared, and the difference in direction numbers is converted into a difference in direction levels. When the difference does not exceed one direction level within two consecutive frames, it is determined to be a state of consistent direction. Only when both the spatial proximity state and the consistent direction state are satisfied simultaneously is it determined that there is a premonition of crosstalk between the target pairs, and the corresponding frame number and the number of the participating target are recorded.

[0108] To avoid misjudgment caused by accidental proximity, the above conditions are not considered as a precursor to crosstalk if they are met only in a single frame, but are only confirmed when they are met in two consecutive frames.

[0109] For target pairs exhibiting signs of potential swapping, their historical entry sequence and edge yielding relationship are read. If one target is in a dominant position in the entry sequence and is marked as a passing target in the edge yielding relationship, then that target is marked as the dominant target; the other target is marked as a subordinate target.

[0110] For subordinate targets, their target number is replaced with the dominant target number in the current frame, and the original number is appended and stored as a spare marker for later recovery. Simultaneously, the association priority of subordinate targets in subsequent frames is reduced; that is, during candidate shrinkage and association, their ranking position is shifted one position to the right. If two targets show no significant difference in their entry sequence or edge yielding relationship, their trajectory convergence trend starting frame number is compared, and the one with the earlier starting frame number is designated as the dominant target. This method achieves a reasonable transfer of target numbers within the intersection area.

[0111] For subordinate targets whose numbers have been replaced, their real-time positions are not directly output in the current frame and the next two frames. Instead, their original position change data is recorded. The center positions of the two consecutive frames before correction are used as reference points, and their average positions are calculated. The average position is obtained by summing the horizontal coordinates of the two center points and dividing by 2, and then performing the same processing on the vertical coordinates. This average position is used as the compensation output position and replaces the original position in the next two frames.

[0112] Simultaneously, the output position of the dominant target is updated normally to ensure the continuity of the overall trajectory. After compensation, the normal output of subordinate targets is restored starting from the 3rd frame, and the spare markers are canceled. Through the above processing, the position jumps caused by target number changes can be reduced, making the final output of multi-target real-time tracking results continuous in time and smooth in space.

[0113] Example 2 illustrates the practical effect of the method in complex dynamic scenarios by combining a specific embodiment. The improvement effect of the technical solution of this application in terms of target identity preservation, substitution suppression and trajectory continuity is verified by comparing experimental data.

[0114] This embodiment selects a bend in an underground mine transport roadway as the test scenario. The video capture frequency is 30 frames per second, the resolution is 1920x1080, and the total test duration is 120 seconds. It includes situations involving mixed traffic of vehicles and personnel, multiple intersections, and dust obstruction. The test data contains a total of 8 targets: 5 vehicles and 3 personnel. There are a total of 14 intersection events and 11 obstruction events. The method of this application is compared with existing conventional methods, which use a direct association method based on location distance and appearance consistency, and operate under the same hardware environment.

[0115] Regarding target tracking accuracy, the target identity retention rate is used as the evaluation metric. The identity retention rate is calculated by summing the total number of frames in which the target identity did not incorrectly switch during the test, dividing by the total number of frames, and then converting the result into a percentage. In this application, the method covers a total of 3600 frames over 120 seconds, with 3384 frames correctly retaining the target identity, corresponding to an identity retention rate of 94%. In the comparative method, the number of frames correctly retaining the target identity is 2876, corresponding to an identity retention rate of 79%. Therefore, it is evident that under complex intersection and occlusion conditions, the method in this application can significantly improve the continuous retention capability of the target identity.

[0116] Regarding target substitution suppression, the number of substitution occurrences is used as the evaluation index. Substitution is determined by the following method: when the same target's number changes in consecutive frames and the original number is not restored, it is counted as one substitution. The proposed method experienced two substitutions during the test, while the comparative method experienced nine. Further analysis shows that in 14 intersection events, the proposed method only experienced a short-term substitution in one complex three-target intersection, while the comparative method experienced substitution in seven of those intersections. This indicates that the preemptive motion corridor and intersection separation fingerprint have a strong ability to distinguish intersection scenarios.

[0117] Regarding trajectory continuity, the trajectory interruption rate is used as the evaluation index. The trajectory interruption rate is calculated by counting the number of interruptions that occurred during target tracking, dividing by the total number of trajectory segments, and expressing the result as a percentage. In this application, the method encountered 5 interruptions across all targets, successfully recovering 4 of them through bidirectional backtracking and reconnection, resulting in a final effective interruption count of 1, with an interruption rate of 3%. In the comparative method, there were 12 interruptions, with no effective reconnection, resulting in a final interruption rate of 18%. This result demonstrates that a reconnection method combining the occlusion boundary dissipation sequence characteristics with turning coherence can effectively recover trajectory breaks caused by occlusion.

[0118] Regarding trajectory smoothness, position jump variables are used as the evaluation index. The position jump variable is calculated by counting the distance between the center positions of two adjacent frames; when this distance is greater than twice the average distance of the previous five frames, it is considered a jump. The method in this application exhibited 3 jumps throughout the entire test, while the comparative method exhibited 15 jumps. Analysis shows that the trajectory transfer correction and output delay compensation in this application can effectively reduce the abrupt position changes caused by changes in target number, resulting in a smoother output trajectory. Further comparative data is provided below: Table 1 Performance Comparison of the Method in this Application and the Comparative Methods Indicator Name This application method Comparison Methods Identity retention rate 94% 79% Number of transpositions 2 9 Track interruption rate 3% 18% Number of position jumps 3 15 As shown in Table 1, in complex dynamic scenarios, the proposed method outperforms existing methods in several key metrics. Especially in situations with frequent target intersections and significant occlusion, the method constrains the candidate range through preemptive motion corridors and achieves exclusive association by combining intersection separation fingerprints, ensuring that targets maintain stable identities even when spatially close. Through bidirectional backtracking, the method enables the recovery of the original trajectory for targets that have been temporarily lost. Furthermore, through trajectory convergence trend identification and transfer correction, potential swapping is adjusted before it occurs, thereby reducing erroneous associations.

[0119] In addition, in terms of real-time performance, under the same processing conditions, the average processing time per frame of the method in this application is 18 milliseconds, while that of the comparative method is 15 milliseconds. Although it is slightly increased, it still meets the requirements of real-time processing and achieves higher tracking stability and accuracy.

[0120] In summary, the method proposed in this application can effectively improve the continuity and reliability of multi-target tracking in complex dynamic scenarios, and has good practical application results.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes, characterized in that, include: Acquire continuous scene images, extract local edge undulation features, micro-displacement direction features and occlusion boundary dissipation order features of each target to form an initial target representation; Based on the initial characterization of the target, a preemptive motion corridor is established for each target, and a convergence separation fingerprint is generated during the target convergence process based on the order of entry, turning offset, and edge yielding relationship. Specifically, based on the initial target characterization, a preemptive motion corridor is established for each target, including: Based on the center position and micro-displacement direction characteristics of the target area, a continuous displacement band is formed by connecting the center positions of the current frame and the previous two frames. The forward extension distance and lateral expansion width are determined according to the micro-displacement intensity to form the initial corridor range. The two sides of the initial corridor range are expanded or contracted according to the local edge undulation characteristics to obtain the corrected corridor. Combining the occlusion boundary dissipation order characteristics, a segmented passage priority order is set at the corridor boundary corresponding to the target contact side to form the preemptive motion corridor. During target convergence, convergence separation fingerprints are generated based on entry sequence, turning offset, and edge yielding relationships. This includes: when preemptive motion corridors overlap, recording the first frame time and entry path position of each target entering the overlapping area to form an entry sequence; statistically analyzing the directional changes in consecutive frames within the convergence area based on micro-displacement direction features to determine turning offset events and their directional shift trends; identifying the sequential changes in edge contraction and expansion of the target contour contact area by combining local edge undulation features and occlusion boundary dissipation order features to form edge yielding relationships; and encoding the entry sequence, turning offset events, and edge yielding relationships in chronological order to form convergence separation fingerprints. The preemptive motion corridor is used to narrow down the candidates for targets at adjacent time points, and the intersection and separation fingerprint is used to perform exclusive association to obtain a unique continuation sequence of targets; For the interrupted segments in the unique continuation sequence, bidirectional backtracking is performed based on the occlusion boundary dissipation order characteristics and the turning continuity before and after the interruption to generate an identity-preserving trajectory. Based on the identity-preserving trajectory, the trajectory convergence trend features within consecutive frames are extracted. When a transponder precursor is detected, trajectory transfer correction and output delay compensation are performed to obtain real-time tracking results for multiple targets.

2. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 1, characterized in that, Acquire continuous scene images, including: Brightness normalization and median filtering are performed on continuous scene images to obtain preprocessed images; foreground candidates are extracted based on the grayscale difference between the current frame and the previous 1 frame and 2 frames, and connected regions are formed after morphological processing; the connected regions are filtered according to area, aspect ratio and boundary integrity to obtain target regions.

3. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 1, characterized in that, Extracting local edge undulation features, including: Each target region is expanded outwards according to the minimum bounding rectangle to form a local analysis range; Extract the outermost closed contour within the local analysis range, and divide the closed contour into 8 equal segments according to length; For each contour segment, the maximum outward distance, the maximum absolute inward distance, the number of distance sign changes, and the average absolute distance are calculated to form local edge undulation features.

4. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 3, characterized in that, Extracting micro-displacement direction features, including: Comparison blocks are placed within each target area, and point-by-point matching is performed around the corresponding position in the previous frame to obtain the micro-displacement of each comparison block; comparison blocks with grayscale variance lower than the preset value or unstable matching are removed, and valid comparison blocks are retained. The micro-displacements of all valid comparison blocks are statistically analyzed by direction to determine the principal direction. The micro-displacement intensity is determined based on the displacement length, and the lateral drift is determined based on the proportion of offset perpendicular to the principal direction, thus forming the micro-displacement direction characteristics.

5. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 4, characterized in that, Extracting occlusion boundary dissipation order features and forming an initial target representation, including: For neighboring target pairs, the contact side is determined and the contact side contour is divided into 4 continuous boundary segments. The boundary point density, inner and outer grayscale contrast and continuous length ratio of each boundary segment are calculated in consecutive frames, and the boundary visibility value is obtained accordingly. The dissipation order is determined according to the order in which the boundary visibility value of each boundary segment decreases to the preset proportion of the average boundary visibility value before occlusion, and the occlusion boundary dissipation order feature is formed by combining the recovery order. The local edge undulation features, micro-displacement direction features, and occlusion boundary dissipation order features are written in a fixed field order to form the initial characterization of the target.

6. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 1, characterized in that, The target unique continuation sequence is obtained, including: The preemptive motion corridor of the previous frame is used to perform spatial projection screening of the target in the current frame, and the direction consistency screening is combined with the micro displacement direction features to form a candidate set. The intersection and separation fingerprints in the candidate set are compared with the intersection and separation fingerprints in the historically confirmed continuation sequence one by one, and the candidate targets with conflicts are eliminated. The remaining candidate targets are sorted according to the distance from their center position to the central axis of the preemptive motion corridor, and the continuing targets are determined by combining the inter-frame interval information and micro-displacement intensity. The continuity of the target is confirmed by continuous frame consistency verification to obtain a unique continuity sequence of the target.

7. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 1, characterized in that, Generate an identity-preserving trajectory, including: The interrupted segments in the unique continuous sequence are identified, and the center position of the target area, the micro-displacement direction features, and the dissipation order features of the occlusion boundary of the effective frames before and after the interruption are recorded. The occlusion boundary dissipation order features of candidate targets within the interrupted segment are compared along the reverse and forward time directions to determine segments with consistent dissipation order. Based on the micro-displacement direction characteristics before and after the interruption, the turning continuity section is screened, and the continuation point is selected within the intersection range of the dissipation sorting consistent section and the turning continuity section. Insert the reconnection point into the target's unique continuation sequence and perform continuous frame consistency verification to generate an identity-preserving trajectory.

8. The multi-target real-time tracking optimization algorithm suitable for complex dynamic scenes according to claim 7, characterized in that, Upon detecting pre-switching warning signs, trajectory handover correction and output delay compensation are performed to obtain real-time multi-target tracking results, including: Based on the changes in the distance between the center positions of the target region and the target area within consecutive frames, the trajectory convergence trend features are extracted. Select all targets with trajectory convergence trends, calculate the spatial overlap between two targets in the corresponding frame, and determine that the two targets are in a spatially close state when their trajectory convergence regions have overlapping pixels in the same frame and the overlapping area is greater than 15% of the area of ​​either target region. Compare the micro-displacement direction features of two targets, convert the difference in direction number into a difference in direction level, and determine that the direction is consistent when the difference does not exceed one direction level in two consecutive frames. Only when both the spatial proximity state and the orientation consistency state are met simultaneously is it determined that there is a premonition of crosstalk between the target pairs, and the corresponding frame number and the number of the participating target are recorded. Based on the historical sequence of entry and the marginal yielding relationship, the dominant and subordinate objectives are determined, and the execution trajectory of the subordinate objectives is transferred and corrected. For subordinate targets that undergo trajectory transfer correction, the average position of the center position of consecutive frames before correction is used for compensation output within a preset number of frames to obtain multi-target real-time tracking results.

Images