Unmanned aerial vehicle image cross-domain migration target identification and deduplication method and system

Abstract

The present application relates to unmanned aerial vehicle image processing and artificial intelligence technical field, specifically to a kind of unmanned aerial vehicle image cross-domain migration target identification and deduplication method and system, obtain video frame gray distribution variation and associated detection area form cross-domain feature, extract center position and moving direction and extend path identification interruption generate motion trajectory, according to trajectory prediction reproduction position and expand search screening candidate target combination gray matching, associated multi-trajectory segment forms continuous trajectory, analysis overlap relationship and deduplicate to obtain cross-domain identification result.The present application constructs target time sequence feature by associating gray variation and detection area, enhances continuous description capability, deduces motion continuation to complete trajectory by combining position direction before and after disappearance and gray trend, adjusts search range according to predicted position and realizes target association by combining spatial proximity and feature consistency, merges dispersed trajectory by fusing trajectory time connection and spatial continuation, reduces repeated identification to improve stability.

Description

Technical Field

[0001] This invention relates to the fields of UAV image processing and artificial intelligence technology, and in particular to a method and system for cross-domain target recognition and deduplication of UAV images. Background Technology

[0002] The field of UAV image processing and artificial intelligence technology includes UAV remote sensing image acquisition, flight path planning, image sequence construction, image preprocessing, target detection, and temporal tracking. Its core lies in acquiring continuous image data through UAVs and performing unified spatiotemporal processing. It performs position association and feature representation on dynamic targets in the images, describes targets by constructing information such as pixel grayscale texture and edges, and establishes motion relationships between consecutive frames. This achieves a complete processing flow from image acquisition and target recognition to trajectory analysis, while also adapting the processing to the differences in different aquatic environments.

[0003] Among them, the method and system for cross-domain migration target recognition and deduplication of UAV imagery refers to the processing solution for target recognition and duplicate counting problems under different water conditions. Its technical aspects include cross-scene sample construction, water emergence behavior time modeling, candidate region extraction, trajectory association, target matching, trajectory stitching, and quantity inference. Specifically, it involves collecting images of multiple water areas and establishing a set of labeled samples, constructing a time window based on the target's water emergence time range, extracting candidate regions in each frame based on grayscale differences, texture changes, and edge contours, and establishing trajectories in consecutive frames based on the displacement direction and area changes of the target's center position. After the target disappears, a search area is set based on the historical position range and time interval for re-matching. At the same time, different time thresholds are set according to whether the target is at the edge or center of the image to retain or terminate the trajectory. In the offline stage, trajectory stitching is completed by spatial overlap ratio, time interval, and feature similarity, and the number of observations and flight path coverage are combined to perform statistical processing on the quantity.

[0004] Current technologies for surface target identification primarily rely on correlation between continuous inter-frame positional changes and target appearance features. When a target experiences brief absences, changes in posture, or deviations in its movement path, historical information is difficult to maintain effectively, easily leading to fragmented target trajectories. For example, when a finless porpoise reappears after diving to the surface, the previous and subsequent detection results may be considered different targets, affecting target count statistics. Furthermore, fixed search ranges and correlation conditions are ill-suited to adapting to changes in target motion. When the target's disappearance time increases or its movement distance expands, the rematching range is prone to deviating from its actual location, resulting in missed detections or mismatches, impacting the accuracy of target identification in cross-water patrol images. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for cross-domain migration target recognition and deduplication in UAV imagery. The technical solution is as follows: A method for cross-domain target recognition and deduplication in UAV imagery includes the following steps: S1: Acquire video frames from the drone, extract grayscale information of the Yangtze finless porpoise detection area, classify pixels according to grayscale levels, extract grayscale distribution changes in continuous video frames, identify continuous areas of grayscale change direction and associate them with the detection area to obtain cross-domain grayscale change features. S2: Based on the cross-domain grayscale change features, extract the change in the center position of the detection area in continuous video frames, identify the target movement direction between adjacent frames, extend the position change path along the movement direction, and combine the target disappearance to form an interruption state and the grayscale change direction to form a motion continuation path, thereby obtaining the target motion extension trajectory. S3: Based on the target motion extension trajectory, identify the target's detection position and movement direction before disappearance, deduce the reappearance position along the movement direction, form a spatial search area centered on the predicted position, gradually expand the search area according to the target's disappearance duration, extract newly appearing detection areas in subsequent video frames, combine spatial positional relationships and grayscale change direction to filter candidate targets, complete target matching, and obtain the candidate association state of the disappeared target. S4: Based on the candidate association state of the disappeared target, extract the target trajectory segments in the cruise video, identify the temporal connection relationship and spatial continuity relationship of the trajectory segments, extend the trajectory position along the movement direction, and filter the associated segments by combining the consistency of the movement direction and the consistency of gray scale changes, and merge them into the same target identifier to obtain the cross-frame continuous target trajectory. S5: Extract target recognition information from the UAV cruise image based on the cross-frame continuous target trajectory, identify the temporal overlap and spatial coincidence relationship between different trajectories, filter duplicate targets by combining grayscale change consistency, remove duplicate target trajectories, and obtain the cross-domain target recognition deduplication result.

[0006] As a further aspect of the present invention, the cross-domain grayscale change features include grayscale level distribution, feature change amplitude, and temporal correlation attributes; the target motion extension trajectory includes trajectory node sequence, motion range boundary, and trajectory direction features; the disappearing target candidate correlation state includes candidate target attributes, correlation credibility, and target correspondence; the cross-frame continuous target trajectory includes trajectory number, target identity attribute, and trajectory continuity marker; and the cross-domain target recognition deduplication result includes target differentiation information, duplicate target set, and valid recognition record.

[0007] As a further aspect of the present invention, the step of obtaining the cross-domain grayscale change feature is as follows: S101: Acquire video frames of the Yangtze River surface captured by the UAV's airborne camera, extract the pixel coordinates and pixel grayscale information corresponding to the Yangtze finless porpoise detection area in each video frame, classify the pixels in the detection area according to the grayscale level, aggregate the pixel positions corresponding to the same grayscale level into a grayscale pixel set, arrange each grayscale pixel set according to the video frame acquisition order, and associate the grayscale level with the detection area identifier to obtain a grayscale pixel distribution matrix. S102: Based on the grayscale pixel distribution matrix, extract the set of pixels corresponding to the same grayscale level according to the video frame arrangement order, compare the distribution range and distribution state of each grayscale level in adjacent video frames, assign change direction markers to the expansion, contraction and stable states respectively, verify whether the direction markers remain in the same direction in consecutive video frames, and arrange the same-direction change markers, corresponding grayscale levels and frame order in association to obtain the grayscale distribution change sequence. S103: Call the grayscale distribution change sequence and the grayscale pixel distribution matrix to map the continuous direction markers to the corresponding video frame pixel positions. Aggregate regions with consistent grayscale change directions based on the spatial connection relationship of adjacent pixels. Filter regions with connected positions and the same direction markers in consecutive video frames. Spatially correspond each region to the Yangtze finless porpoise detection region in the same frame. Collect the detection region identifier, grayscale level, change direction and frame sequence information to obtain cross-domain grayscale change features.

[0008] As a further aspect of the present invention, the step of obtaining the target motion extension trajectory is as follows: S201: Based on the cross-domain grayscale change features, extract the boundary coordinates of the Yangtze finless porpoise detection area in continuous video frames. Use the median of the horizontal coordinates of the left and right boundaries as the center horizontal coordinate, and the median of the vertical coordinates of the upper and lower boundaries as the center vertical coordinate. Arrange the center coordinates according to the frame order, compare the signs of the center coordinate differences of adjacent frames in the horizontal and vertical directions of the image, and collect the corresponding direction signs, frame order, region identifiers and center coordinates to obtain the center displacement vector set. S202: Call the center displacement vector set and the cross-domain grayscale change feature, search the detection area corresponding to the same target identifier along the frame sequence, identify the time period when the target identifier does not appear in adjacent frames and reappears in subsequent frames, extract the end center coordinates, end movement direction, grayscale change direction and missing frame segments before disappearance, compare the spatial orientation of the center coordinates of the reproduced area relative to the end movement direction, merge the same orientation information with the missing frame segments, and obtain the target interruption association table; S203: According to the target interruption association table, call the center displacement vector set, continue the detection area change path along the movement direction of the end before disappearance, arrange the missing frame segment path nodes according to the change trend of adjacent center coordinates before disappearance, filter the target that is first in the spatial distance sorted with the end of the continued path and has the same gray change direction, and splice the path before disappearance, the missing frame segment path nodes and the reconstructed path according to the frame order, and use the same target identifier to obtain the target motion extension trajectory.

[0009] As a further aspect of the present invention, the step of obtaining the candidate association state of the disappeared target is as follows: S301: Based on the target motion extension trajectory, extract the end detection position before disappearance, the end movement direction and the adjacent trajectory nodes of the end, compare the relationship between the displacement direction of the adjacent trajectory nodes and the end movement direction, translate the position change between nodes with the same direction frame by frame along the end movement direction to the subsequent frames, arrange the predicted positions according to the frame order, and collect the target identifier, interruption start frame, end detection position, movement direction and the predicted positions of each frame to obtain the reconstructed position prediction domain; S302: Call the reproducible position prediction domain, expand the search boundary around each inferred position frame by frame according to the number of frames the target identifier interrupts, extract the detection area that first enters the corresponding search boundary in subsequent video frames and does not use the original target identifier, filter the first region of the spatial interval between the center position of the detection area in each frame and the inferred position in the same frame, and collect the region identifier, center position, appearance frame order and search boundary to obtain the reproducible target candidate set; S303: Based on the proposed set of reproducible target candidates, the cross-domain grayscale change features and the reproducible position prediction domain are called to compare the grayscale change direction of each candidate region with the grayscale change direction of the target before disappearance. Candidate regions with the same direction are associated with the original target identifier. The target identifier, prediction position, reproducible position, interrupted frame segment, candidate region identifier and grayscale change direction are merged in frame order to obtain the disappearing target candidate association state.

[0010] As a further aspect of the present invention, the step of obtaining the cross-frame continuous target trajectory is as follows: S401: Based on the candidate association state of the disappeared target, extract the start frame, end frame, first position, last position, movement direction and grayscale change direction of each target trajectory segment in the cruise video. Arrange the trajectory segments according to the start frame. Compare the frame order of the end frame of the previous segment with the start frame of the next segment. Mark the cases of adjacent frame order and interruption intervals respectively. Extend the last position of the previous segment frame by frame according to the interrupted frame segment and correspond it with the first position of the next segment to obtain the trajectory connection relationship table. S402: Call the trajectory connection relationship table and the candidate association state of the disappearing target, extract the end movement direction, the beginning movement direction, the position extension direction and the gray change direction of each connection segment, filter the trajectory segments with the same movement direction, the beginning position of the next segment is in the extension direction of the previous segment and the gray change direction is consistent, and classify the previous segment identifier, the next segment identifier, the connection frame segment and the direction corresponding state into the same association unit to obtain the continuous segment association chain; S403: Based on the continuous segment association chain, call the trajectory connection relationship table, arrange the trajectory segments in the same association chain according to the segment start frame and end frame, connect the end position of the previous segment, the interrupted frame segment and the beginning position of the next segment in sequence, filter the association chain where the movement direction and grayscale change direction of adjacent segments are consistent, replace the target identifier of the first segment of the starting frame with the identifier of the other segments, collect the position and direction information of each frame, and obtain the cross-frame continuous target trajectory.

[0011] As a further aspect of the present invention, the step of obtaining the cross-domain target recognition deduplication result is as follows: S501: Extract target identification information from the UAV's cross-water cruise image based on the cross-frame continuous target trajectory, and collect the trajectory start and end frames, intra-frame detection areas, movement direction, grayscale change direction and water index according to the target identifier. Compare the frame sequence intersection of different trajectories with the overlap of detection areas in the same frame, and arrange the trajectories that have a common frame sequence and whose detection areas overlap to obtain a trajectory spatiotemporal overlap map. S502: Call the aforementioned trajectory spatiotemporal overlap map, extract the center position of the detection area, the area boundary and the target identifier within the common frame segment of each trajectory combination, filter the trajectory combinations whose center position falls into the boundary of the other's area and whose target identifiers are different, and group the trajectory identifier, common frame segment, overlapping area, movement direction, grayscale change direction and water area index into the same data unit to obtain the duplicate target pointing cluster; S503: Based on the repeated target pointing cluster, the trajectory spatiotemporal overlap map is called to compare the gray-scale change direction, movement direction and common frame segment position change relationship of each trajectory in the same data unit. The trajectory identifiers that are continuous before and after the common frame segment and have records of start and end frames, intra-frame detection area, movement direction, gray-scale change direction and water area index are retained. The detection areas and frame order of the remaining trajectories are merged into the retained identifiers. The repeated target identifiers are deleted to obtain the cross-domain target recognition deduplication result.

[0012] A system for cross-domain target recognition and deduplication in UAV imagery, the system comprising: The grayscale migration module acquires video frames of water patrol captured by the UAV's onboard camera, extracts the pixel grayscale information of the Yangtze finless porpoise detection area, classifies the pixel distribution according to grayscale level, extracts the direction of grayscale distribution change in continuous video frames, and associates the continuous area of ​​grayscale change direction with the detection area to obtain cross-domain grayscale change features. The trajectory extension module extracts the center position change information of the finless porpoise detection area in continuous video frames based on the cross-domain grayscale change features, extracts the target movement direction based on the position change direction of adjacent video frames, extends the change path of the detection area along the movement direction, identifies the intermittent disappearance state of the target in the video frame, and obtains the target movement extension trajectory by combining the position change relationship before and after disappearance with the grayscale change direction. The target retrieval module extracts the target's detection position and movement direction before disappearance based on the target's motion extension trajectory, deduces the target's reappearance position along the movement direction, forms a search area centered on the predicted position, expands the search area range based on the change in the target's disappearance duration, extracts the center position of the Yangtze finless porpoise detection area in subsequent video frames, compares the spatial relationship between the center position of the detection area and the predicted position, filters the position change-related areas, and completes target matching by combining the grayscale change direction to obtain the candidate related state of the disappeared target. The trajectory merging module extracts target trajectory segments from the UAV cruise video based on the candidate association states of the disappeared target, filters associated segments according to the temporal connection and spatial continuity relationship between trajectory segments, connects continuous trajectory positions along the movement direction, filters associated trajectory segments by combining the movement direction change and grayscale change direction, merges the target identifiers corresponding to multiple trajectory segments, and obtains cross-frame continuous target trajectories. The deduplication identification module extracts target identification information from the cross-waterway cruise image based on the cross-frame continuous target trajectory, filters duplicate target trajectories according to the temporal and spatial relationships of different trajectories, extracts duplicate pointing targets by combining the grayscale change direction, removes duplicate target information, and obtains the cross-domain target identification deduplication result.

[0013] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, by associating the grayscale distribution changes in continuous video frames with the detection area, a feature representation relationship of the target changing over time is constructed, enabling the target state under water surface disturbance to have the ability to continuously change. By combining the position changes, movement direction, and grayscale change trends before and after the target disappears to deduce the motion continuity relationship, the target motion information in the short-term missing state is supplemented, reducing the impact of trajectory interruption on target association. By adjusting the spatial search range based on the predicted position and combining spatial proximity and feature consistency for target association, the possibility of incorrect correspondence when the target reappears is reduced. By fusing the temporal connection, spatial continuity, and feature consistency relationship of trajectory segments, scattered trajectories are merged into a unified target identifier, reducing repeated identification during cross-regional patrols and improving the stability of target statistical results. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the acquisition of cross-domain grayscale variation features according to the present invention. Figure 3 This is a flowchart illustrating the process of obtaining the target motion extension trajectory in this invention. Figure 4 This is a flowchart illustrating the process of obtaining the candidate association state of the disappeared target in this invention. Figure 5This is a flowchart of the process for obtaining continuous target trajectories across frames according to the present invention; Figure 6 This is a flowchart illustrating the process of obtaining the cross-domain target recognition and deduplication results of this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1 to 6 A method for cross-domain target recognition and deduplication in UAV imagery includes the following steps: S1: Acquire video frames of the Yangtze River surface captured by the UAV's onboard camera, extract the grayscale information of the corresponding image pixels in the Yangtze finless porpoise detection area, classify the pixels according to the grayscale level, extract the distribution changes of each grayscale level in continuous video frames, identify the areas where the grayscale change direction is continuous and associate them with the corresponding detection areas to obtain cross-domain grayscale change features. S2: Based on cross-domain grayscale change features, extract the change of the center position of the Yangtze finless porpoise detection area in continuous video frames, identify the movement direction of the detection area between adjacent video frames, extend the change path of the detection area along the movement direction, identify the intermittent disappearance of the target in continuous video frames to form an interruption of the marker, combine the position change relationship before and after disappearance with the grayscale change direction to form the target motion continuation path, and obtain the target motion extension trajectory. S3: Identify the detection position and movement direction of the target before it disappears based on the extended trajectory of the target's motion. Infer the position where the target will reappear along the movement direction. Form a spatial search area centered on the predicted position. Gradually expand the range of the spatial search area based on the extension of the target identifier interruption duration. Extract newly appearing detection areas in subsequent video frames. Filter candidate targets based on the spatial proximity between the newly appearing detection areas and the predicted positions. Combine the consistency of grayscale change direction to complete target matching and obtain the candidate association state of the disappeared target. S4: Extract multiple target trajectory segments from the cruise video based on the candidate correlation state of the disappeared target, identify the temporal connection relationship and spatial continuity relationship between the trajectory segments, extend the position of the next trajectory along the movement direction of the previous trajectory, identify trajectory segments with continuous position changes, filter associated trajectory segments by combining the consistency of trajectory movement direction and the consistency of gray scale change, and merge multiple trajectory segments that meet the continuity relationship into the same target identifier to obtain cross-frame continuous target trajectories. S5: Extract target recognition information from UAV cross-water cruise images based on continuous target trajectories across frames, identify the temporal overlap and spatial coincidence relationships between different trajectories, filter out target trajectories with repeated directions, and remove duplicate targets by combining grayscale change consistency to obtain cross-domain target recognition and deduplication results.

[0017] Cross-domain grayscale change features include grayscale level distribution, feature change amplitude, and temporal correlation attributes; target motion extension trajectory includes trajectory node sequence, motion range boundary, and trajectory direction features; disappearing target candidate correlation status includes candidate target attributes, correlation credibility, and target correspondence; cross-frame continuous target trajectory includes trajectory number, target identity attribute, and trajectory continuity marker; cross-domain target recognition deduplication results include target differentiation information, duplicate target set, and valid recognition record.

[0018] Please see Figure 2 The steps for obtaining cross-domain grayscale variation features are as follows: S101: Acquire video frames of the Yangtze River surface captured by the UAV's airborne camera, extract the pixel coordinates and pixel grayscale information corresponding to the Yangtze finless porpoise detection area in each video frame, classify the pixels in the detection area according to the grayscale level, aggregate the pixel positions corresponding to the same grayscale level into a grayscale pixel set, arrange each grayscale pixel set according to the video frame acquisition order, and associate the grayscale level with the detection area identifier to obtain a grayscale pixel distribution matrix. For the execution phase of acquiring video frames of the Yangtze River surface and subsequent feature acquisition using an airborne camera on a UAV, the UAV was configured to cruise at a constant speed over a fixed monitoring area of ​​the Yangtze River. The camera's resolution was preset to 1920 x 1080, and the sampling frame rate was configured to 30 frames per second. Image data from frames 1 to 30 of the video sequence was extracted. The detection region for finless porpoises within each frame was located using a bounding box, and the coordinates of the four vertices of the corresponding rectangular bounding box were obtained. Taking frame 1 as an example, the bounding box coordinates of the finless porpoise detection region identified as ID_1 were read as 200, 300, 250, and 320. The two-dimensional pixel coordinates and their single-channel grayscale values, limited to integers between 0 and 255, were extracted for each pixel. To classify and group pixels within the detection region according to their grayscale level, an experiment was conducted using 100,000 pixel samples containing various media. The baseline thresholds for grayscale level classification were set to 80 and 160. Based on the aforementioned baseline thresholds, pixels within the range of 0 to 80 are classified as Level_1, those between 81 and 160 as Level_2, and those between 161 and 255 as Level_3. For example, for pixels with coordinates of 210 and 310 and a grayscale value of 65 within the detection area of ​​the first frame, their coordinate data is appended to the corresponding Level_1 set. After traversal, spatial set merging logic is performed on the pixel position coordinates corresponding to the same grayscale level based on the grayscale level identifier, generating mutually independent grayscale pixel sets. Then, according to the order of video frame timestamp acquisition, a key-value pair mapping relationship is established between each set and a globally unique identifier to obtain a multi-level grayscale pixel distribution matrix.

[0019] S102: Based on the gray-level pixel distribution matrix, extract the set of pixels corresponding to the same gray level according to the video frame arrangement order, compare the distribution range and distribution state of each gray level in adjacent video frames, assign change direction labels to the expansion, contraction and stable states respectively, verify whether the direction labels remain in the same direction in consecutive video frames, and arrange the same-direction change labels, corresponding gray levels and frame order in association to obtain the gray-level distribution change sequence. The grayscale pixel distribution matrix is ​​retrieved, and the internal key-value pair indices are read. Pixel sets with the same grayscale level label within adjacent video frames are extracted in pairs, strictly according to the video frame timestamp order. Taking the Level_1 grayscale level, identified as ID_1, as an example, the grayscale pixel sets corresponding to the first and second frames are extracted, with the total number of pixels being 450 and 480 respectively. Based on the difference in the total number of pixels of the same level within two adjacent frames and the comparison logic with the baseline number, the area change rate is obtained to be approximately 0.0667. A water surface disturbance diffusion calibration test database containing 100 samples is called, and the state division judgment threshold is set to ±0.05. The obtained area change rate value is substituted into the threshold judgment condition. Since the measured value of 0.0667 is greater than the baseline threshold of 0.05, a change direction marker Exp representing the expansion state is written to this data unit. The pixel sets corresponding to the second and third frames are extracted for further verification. With a measured number of 515 pixels, the change rate is approximately 0.0729, and the Exp marker is continuously assigned. During the temporal comparison process, the direction marker is kept constant for 5 consecutive video frames. The verified direction marker maintains the same direction attribute. Then, the changed direction marker Exp that has passed the same direction verification, its corresponding gray level Level_1, and the consecutive start and end frame sequences involved in the verification are combined into a nested array structure and stored in associated memory. This effectively filters out single-frame gray level change noise and obtains the gray level distribution change sequence.

[0020] S103: Call the grayscale distribution change sequence and grayscale pixel distribution matrix, map the continuous direction markers to the corresponding video frame pixel positions, aggregate regions with consistent grayscale change directions based on the spatial connection relationship of adjacent pixels, filter regions with connected positions and the same direction markers in consecutive video frames, spatially correspond each region to the Yangtze finless porpoise detection region in the same frame, and collect the detection region identifier, grayscale level, change direction and frame sequence information to obtain cross-domain grayscale change features.

[0021] The grayscale distribution change sequence and the grayscale pixel distribution matrix containing discrete coordinates are invoked to extract the approved continuous direction marker Exp and its corresponding frame sequence interval recorded in the associated data group. This continuous direction marker is then backfilled to the specific pixel position in the corresponding video frame's coordinate system using a coordinate system mapping method. Taking the newly added extended pixels as the processing object, an 8-neighborhood spatial connectivity verification logic is performed on each pixel. If the pixel spatial distance is less than or equal to the judgment limit, it is determined that the adjacent pixel spatial connectivity relationship is satisfied and they converge into a connected region block. Based on the geometric mass distribution model, the centroid coordinates of the aggregated connected region blocks within each video frame are obtained, and the spatial displacement of the centroid position is measured between consecutive video frames. For example, the measured centroid distance between the 2nd and 3rd frames is approximately 3.605 pixels. Combining the drone hovering shooting height and the maximum swimming speed of the Yangtze finless porpoise, the judgment benchmark value is clearly defined as 15 pixel units. Since the measured distance does not exceed the benchmark value, the grayscale change area between consecutive frames is determined to meet the positional connection standard. Simultaneously, it compares whether these regions hold completely consistent Exp direction markers in the sequence library, confirms that the centroid coordinates fall within the corresponding bounding box interval through the boundary coordinates including verification logic, and finally creates a comprehensive data feature vector and adds relevant identifiers and frame sequence numbers to obtain cross-domain grayscale change features.

[0022] Please see Figure 3 The steps for obtaining the target's extended trajectory are as follows: S201: Based on cross-domain grayscale variation features, extract the boundary coordinates of the Yangtze finless porpoise detection region in continuous video frames. Use the median of the horizontal coordinates of the left and right boundaries as the center horizontal coordinate, and the median of the vertical coordinates of the upper and lower boundaries as the center vertical coordinate. Arrange the center coordinates according to the frame order, compare the signs of the center coordinate differences between adjacent frames on the horizontal and vertical axes of the image, and collect the corresponding signs of the direction signs, frame order, region identifiers and center coordinates to obtain the center displacement vector set. Based on the cross-domain grayscale variation feature data table structure, iterative extraction operations are performed on consecutive video frames storing specific target identification information. The coordinates of the four corners of the bounding box of the finless porpoise detection region within the current video frame's image coordinate system are retrieved, and the center feature point of the bounding box is obtained based on the diagonal cross-location rule. Taking the bounding box boundary value of the first frame with the identifier ID_1 as an example, its center x-coordinate is located as 225 and its center y-coordinate as 310, which are combined into the center coordinates of that frame. Following the video frame order and auto-incrementing mechanism, the center x-coordinate of the bounding box corresponding to the next adjacent frame, i.e., the second frame, is obtained as 230.5 and its y-coordinate as 313.5. Based on the spatial difference analysis of the center coordinates of adjacent two frames, the positional changes on the two orthogonal components of the image coordinate axes are obtained: the coordinate change value along the horizontal axis is 5.5, and the coordinate change value along the vertical axis is 3.5. A symbolic judgment logic is introduced; since the changes in the horizontal and vertical axes both show positive increments, they are assigned a positive sign of +1 for both directions. Create a sequence set record table with a multi-field structure, and fill the table with the horizontal and vertical direction symbols obtained by the judgment, the current comparison frame sequence involved in the analysis, the unique identifier ID_1 of the detection area involved, and the measured values ​​of the center coordinates involved, to obtain the center displacement vector set.

[0023] S202: Call the center displacement vector set and cross-domain grayscale change features, search the detection area corresponding to the same target identifier along the frame sequence, identify the time period when the target identifier does not appear in adjacent frames and reappears in subsequent frames, extract the end center coordinates, end movement direction, grayscale change direction and missing frame segments before disappearance, compare the spatial orientation of the center coordinates of the reproduced area relative to the end movement direction, merge the same orientation information with the missing frame segments, and obtain the target interruption association table; The system retrieves the center displacement vector set and cross-domain grayscale change feature dataset items, and initiates a forward traversal search based on the video frame sequence timeline to locate the historical detection region sequence records corresponding to the same target identifier. During continuous frame inspection, if a continuous blank frame sequence containing a recording interruption is detected, and subsequently a region with a newly generated initial identifier but matching the finless porpoise morphology reappears, then this complete interval is marked as the target interruption period. For example, if ID_1 was found to have no corresponding region record from the last coordinate record in frame 15 to frame 45, and a candidate detection region box is generated again in frame 46, the system extracts the center coordinates at the end of frame 15 and extracts the previous direction features to obtain the end movement direction vector as horizontal +4 and vertical +2. Simultaneously, it retrieves the grayscale change direction flag bit Exp_w recorded in the feature library for this period, and extracts the difference in frame sequence span between the first and last frames to obtain a total of 30 missing frames. For the center coordinates of the candidate region that reappears in frame 46, the spatial vector of the reconstructed displacement relative to the vanishing point is extracted. Then, using a cosine similarity calculation logic in a Cartesian coordinate system, the directional similarity between the reconstructed relative spatial orientation vector and the final movement direction vector is obtained, which is approximately 0.9976. This is compared with the same-direction orientation discrimination coefficient threshold of 0.85 set by actual measurements of the finless porpoise's swimming trajectory segment. It is determined that the spatial location of the new target is within a reasonably extended sector of the original direction. Subsequently, the determined state information and the missing frame segment are imported into a multi-row data linked list for storage, resulting in a target interruption association table.

[0024] S203: Based on the target interruption association table, call the center displacement vector set, continue the detection area change path along the direction of movement of the end before disappearance, arrange the missing frame segment path nodes according to the trend of adjacent center coordinate changes before disappearance, filter the target that is first in spatial distance from the end of the continued path and has the same grayscale change direction, and splice the path before disappearance, the missing frame segment path nodes and the reconstructed path in frame order, and use the same target identifier to obtain the target motion extension trajectory.

[0025] Based on the data index pointer provided by the target interruption association table, the center displacement vector set data module generated in the previous stage is called to start the breakpoint trajectory continuation program. First, the center coordinates of the end record before disappearance and the constant movement direction vector of the end are loaded. Using the time missing span as an extrapolation variable, the change path during the invisible period is inferred based on the assumption of constant displacement in a single frame. The theoretical position of the last node displacement corresponding to the 46th frame is deduced to be coordinates 420 and 460. The focus is turned to several candidate areas of water surface detection targets that are re-presented in the 46th frame. For each reproduced target and its corresponding continuation path end, the two-dimensional plane space Euclidean distance solution logic is performed. The spatial distance of target ID_X is obtained as 10 pixels, and the spatial distance of interference target ID_Y is about 36.05 pixels. After forward bubble sorting of the obtained distance values, the ID_X corresponding to the first target with the shortest distance is extracted. The initial grayscale change direction identifier record is verified to be the extended identifier Exp_w corresponding to the water effluent physical logic. It is determined that it has the same grayscale change indication property as the specific state record before disappearance. Subsequently, a complete chain connection operation of the path data structure is performed, combining the measured path record, the deduced node record, and the reproduced path record. The original identifier ID_X of the candidate is removed, and the original historical target identifier ID_1 is used globally to overwrite the attribute value, thus obtaining the extended trajectory of the target motion.

[0026] Please see Figure 4 The steps for obtaining the candidate associated state of the disappeared target are as follows: S301: Based on the target's motion trajectory, extract the end detection position before disappearance, the end movement direction, and adjacent trajectory nodes. Compare the relationship between the displacement direction of adjacent trajectory nodes and the end movement direction. Translate the position changes between nodes with the same direction frame by frame along the end movement direction to subsequent frames. Arrange the predicted positions according to the frame order. Collect the target identifier, interruption start frame, end detection position, movement direction, and predicted positions of each frame to obtain the reconstructed position prediction domain. Using the target motion trajectory data set as a baseline reference, for target ID_1 with a recorded hidden fault, the end detection position of the last frame before entering a continuous invisible state, i.e., frame 15, is extracted. Trajectory segments are retrieved backward to obtain the forward adjacent trajectory node of this node, i.e., the center coordinate of frame 14. Based on the local coordinate difference, the local instantaneous displacement pointing vector is obtained as horizontal +4 and vertical +2. A numerical comparison and matching operation is performed between this local node displacement pointing vector and the overall vector of the end movement direction, determining that the corresponding orthogonal component values ​​of the two vectors are perfectly consistent. A translation operation baseline framework is established, using the end detection position as the starting point, and position coordinate mapping is performed to each missing subsequent frame according to the increment of the end movement direction. For frame 16, a horizontal cumulative offset of 4 and a vertical cumulative offset of 2 are assigned to obtain its calculated position. This position state recursion process continues until the calculated position of the cumulative horizontal coordinate 424 and cumulative vertical coordinate 462 in frame 46 is obtained. The obtained inference location set is arranged into an ordered array list according to the sequence number of each video frame involved. The target identifier ID_1 of the interruption, the interruption start frame number, the end detection position, the overall movement direction scalar, and the sequence number of each independent frame in the interruption area and the corresponding frame-by-frame inference location information are written into the newly created multi-mapping association form object to obtain the reproducible location prediction domain.

[0027] S302: Call the reproducibility location prediction domain, expand the search boundary around each inferred location frame by frame according to the number of frames the target identifier interrupts, extract the detection area that first enters the corresponding search boundary in subsequent video frames and does not use the original target identifier, filter the first region in the spatial interval between the center position of the detection area in each frame and the inferred position in the same frame, and collect the region identifier, center position, appearance frame order and search boundary to obtain the reproducibility target candidate set; The system calls upon the reconstructed location prediction dataset and, for the specific inferred coordinates stored therein, retrieves the parameter value of the number of frames of interruption that have occurred since the interruption occurred, as a variable factor. For the 46th frame data slice, which has experienced an interruption duration of 30 frames, the inferred location point of this frame is extracted as the center of the planar two-dimensional geometric search circle. The dynamic external amplification amount is determined based on the interruption frame duration. Combining the pre-calibrated average deviation expansion coefficient of 1.5 pixels per frame and the interruption duration, a spatial search boundary radius of 45 pixels is obtained. Based on this search boundary radius, a covered circular area is delineated, and the system then performs target boundary detection and extraction operations on the same image segment, screening for objects that first enter the currently predetermined search boundary circular area. Assuming the intercepted intrusion areas Area_U and Area_V are identified, the spatial straight-line interval parameter between their center positions and the corresponding inferred theoretical positions in frame 46 is obtained based on the spatial two-point distance measurement rule. The spatial interval for Area_U is approximately 12.649 pixels, and the spatial interval for Area_V is approximately 40.249 pixels. After comparison, the region Area_U with the smallest numerical interval parameter is selected as the first region in the sorting list and extracted and stored. Its new temporary region identification information, center coordinate data, appearance frame order and corresponding dynamically calculated search boundary radius are packaged and managed to obtain the candidate set of reproducible targets.

[0028] S303: Based on the candidate set of reproducible targets, call the cross-domain gray-scale change features and the prediction domain of the reproducible position, compare the gray-scale change direction of each candidate region with the gray-scale change direction of the target before disappearance, associate the candidate regions with the same direction with the original target identifier, and merge the target identifier, prediction position, reproducible position, interrupted frame segment, candidate region identifier and gray-scale change direction in frame order to obtain the disappearing target candidate association state.

[0029] Based on the target candidate set data structure, the system collaboratively calls the pre-generated cross-domain grayscale change feature list and the newly established reconstructed location prediction domain data group from the system library. The first candidate region object extracted from the candidate set is selected, and the corresponding cross-domain grayscale feature item marker record representing water surface expansion is extracted when the detected region changes. A reverse data call request is executed to query the file corresponding to the original target before disappearance, mentioning that the grayscale change direction index item record of the water surface disturbance diffusion before its disappearance is also a specific state marker Exp_w with water ripple associated with it. In the direction indicator value matching module, the content of the extracted two marker strings is compared for consistency, determining that the candidate region is completely consistent with the last grayscale evolution interaction state presented before complete disappearance. Based on this consistency determination instruction, the newly generated identifier attached to the candidate object is hard-associated and bound to the earliest recorded original target identifier name through memory pointer redirection, and equivalent identifier assimilation processing is performed. In the data aggregation stage, the data volume containing the original target identifier, predicted location coordinates, detected and measured coordinates, recorded missing blank span intervals, early temporary identifiers of the assimilated candidate areas, and grayscale change direction identifiers are packaged and integrated, and then linearly rearranged and merged according to the sequence number of the video frames from the start to the end covered, to obtain the disappearing target candidate association state.

[0030] Please see Figure 5 The steps for obtaining continuous target trajectories across frames are as follows: S401: Based on the candidate association state of the disappeared target, extract the start frame, end frame, first position, last position, movement direction and grayscale change direction of each target trajectory segment in the cruise video. Arrange the trajectory segments according to the start frame, compare the frame order of the end frame of the previous segment with the start frame of the next segment, mark the adjacent frame order and the cases with interruption intervals respectively, extend the last position of the previous segment frame by frame according to the interrupted frame segment, and correspond it with the first position of the next segment to obtain the trajectory connection relationship table. Based on the candidate correlation state mapping dictionary file of the disappeared target, a source tracing and capture operation is performed to comprehensively extract all captured scattered or independent trajectory segment data blocks. Each segment unit is unpacked to extract key parameters. All segment data blocks are sorted in ascending order based on the extracted first-end occurrence frame time sequence as the sorting key. The final termination frame sequence value (15) of the trajectory segment at the earlier time position and the first starting frame sequence value (46) of the trajectory segment in the next order are extracted sequentially. The frame number gap value difference is obtained by looking up the first and last frame numbers in a table and finding it to be 31 frames. A classification logic judgment condition is set: since the value difference is significantly greater than 1, it indicates that there is a gap time period without entity capture observation data. A "Break" tag representing an interruption interval is then added to the correlation information area. The missing frame number 30 in the interrupted frame segment is extracted as the control duration. The position coordinates of the end point of the preceding segment are extracted. Using the cumulative interval of the interrupted frame segment as the gain control variable, the spatiotemporal position extension mapping is performed by successively crossing the same movement direction scalar as the preceding segment to calculate the corresponding virtual spatiotemporal plane of the 46th frame, and the coordinates of the extended prediction azimuth point are obtained. A bidirectional corresponding data attachment chain node is established, and the virtual extension position data is combined and bound with the actual appearance position of the starting end of the target entity segment that follows it to obtain the trajectory connection relationship table.

[0031] S402: Call the trajectory connection relationship table and the candidate association state of the disappearing target, extract the end movement direction, the beginning movement direction, the position extension direction and the gray change direction of each connection segment, filter the trajectory segments with the same movement direction, the beginning position of the next segment is in the extension direction of the previous segment and the gray change direction is consistent, and classify the previous segment identifier, the next segment identifier, the connection frame segment and the direction corresponding state into the same association unit to obtain the continuous segment association chain; The system invokes the linked list structure of the trajectory connection relationship table and the index of candidate associated data for disappearing targets, traversing each record to extract the motion vector sets attached to the connection segments between each pair of groups. It retrieves the measured values ​​of the movement direction vectors at the end of the near-disappearance phase of the preceding character within the group, the initial movement direction vector values ​​of the first segment at the initial water-breaking phase of the following character, and the absolute position extension direction span vector. Simultaneously, it exports the corresponding grayscale change direction indicators such as water area fluctuation change indicators. For the obtained displacement direction parameters of the two segments, a direction consistency assessment logic is introduced to obtain a comparison value of approximately 0.988 for the same direction of movement before and after. Using 80 batches of finless porpoise cruise heading sampling data, a permissible benchmark threshold of 0.8 is set. After verification, 0.988 is greater than the 0.8 threshold, successfully passing the judgment. Following the same logic, the direction conformity between the preceding segment's movement direction and the extended horizontal crossing movement direction vector is obtained as 1.0. Similarly, it is verified that the actual dwelling point position at the beginning of the following segment accurately falls within the inertial inference extension direction of the preceding segment. The grayscale diffusion identifiers of both parties at each key connection period are retrieved, and the Exp_w identifiers are maintained to indicate that the change direction is consistent. After compliance is achieved through the above-mentioned geometric and texture logic direction comparison, the preceding fragment identifier, the following fragment identifier, the missing frame number index, and all overlap verification status information are extracted and injected into the same attribute frame with defined association logic for memory aggregation, resulting in a continuous fragment association chain.

[0032] S403: Based on the continuous segment association chain, call the trajectory connection relationship table, arrange the trajectory segments in the same association chain according to the segment start frame and end frame, connect the end position of the previous segment, the interrupted frame segment and the beginning position of the next segment in sequence, filter the association chain where the movement direction and grayscale change direction of adjacent segments are consistent, replace the target identifier of the first segment of the starting frame with the identifier of the other segments, collect the position and direction information of each frame, and obtain the cross-frame continuous target trajectory.

[0033] Following the successfully constructed, logically guided continuous segment association chain data dictionary, the system reversely calls the fragmented motion record entries archived within the trajectory connection relationship table. For independent trajectory segments encompassed within the same highly correlated association chain container structure, they are arranged one by one in ascending order from early to late time occurrence. The last coordinate point information recorded by the first segment, which is already at the head of the queue and sequentially arranged, is retrieved. Subsequently, based on the progressive coordinate sequence generated by the trajectory smooth transition logic, discrete point data used to fill in the positional data during the period without video capture and recording is connected. Then, the starting coordinate position of the first time of the subsequent segment is connected, forming a network chain connecting multiple nodes of displacement. A global screening and verification process is carried out, checking and traversing the corresponding physical displacement forward direction change rate value and the grayscale change indicator code reflected by the water surface node by node along the entire combined link line. After confirming that all records within the connecting transition zones and extended sections of each sub-segment are highly consistent and unchanging, the entire associated chain is determined to be a chain that satisfies the global unidirectional physical consistency rule and is allowed to pass through and be retained. Using the original, fixed, assigned identifier code of the first trajectory segment appearing at the very beginning of the entire associated composite long chain as the base unique wildcard, a batch replacement and overwrite procedure is executed. This forcibly cleans up all the scattered identifier codes held by subsequent attached sub-trajectory segments and rewrites them to the initial base identifier code. Furthermore, the entire time-period frame-by-frame spatial point coordinate data and the set of attribute vectors along the line are integrated and stored in the persistent memory layer to obtain the cross-frame continuous target trajectory.

[0034] Please see Figure 6 The steps for obtaining the deduplication results of cross-domain target recognition are as follows: S501: Extract target recognition information from UAV cross-water cruise images based on cross-frame continuous target trajectories, and collect trajectory start and end frames, intra-frame detection areas, movement directions, grayscale change directions, and water indexes according to target identifiers. Compare the overlap of frame order intersections and detection areas in the same frame for different trajectories, and combine and arrange trajectories with common frame order and overlapping detection areas to obtain a trajectory spatiotemporal overlap map. Based on the cross-frame continuous target trajectory file library generated by prior logic integration, target recognition and auxiliary detection result data corresponding to the UAV cross-water surface patrol detection video archive image are extracted. From the complex and multi-dimensional information set, related sub-class data are summarized and absorbed according to the global target identifier classification code. The parameter dimensions include the global absolute time start and end frame sequence values, the coordinate group of the two-dimensional rectangular border of the specific detection area, the direction of the two-dimensional translation vector, the grayscale change indicator, and the water area index number. Using the same method, the corresponding detection interval and related frame set parameters are obtained from another segment of identification labels with cross-water shooting numbers. Two trajectory data sets suspected of being redundant due to cross-location co-shooting are retrieved and sent to the frame sequence spatiotemporal comparison module for set intersection logic extraction processing. It is found that the overlap covers a common recording period of up to 51 recorded frames, namely frames 150 to 200. The boundary parameters of two positioning frame groups belonging to different identifiers are extracted at frame 160. Based on the geometric topological relationship measurement rule of the overlapping and overlapping coverage of the two rectangular primitive areas, the intersection-union ratio parameter value is obtained to be approximately 0.429. A dual-camera field-of-view calibration and entanglement verification detection method was introduced, with a proposed threshold of 0.35 as the lower limit for the overlap critical parameter. Measured values ​​exceeding this threshold were verified across different levels, triggering a warning state for spatial intersection and fusion of co-located frames. Two involved trajectory combinations, identified through investigation as having shared corresponding frame sequence time intervals and having accurately verified overlapping and fusion states in the frame's physical cross-section, were paired, bundled, and loaded into a single data unit package to obtain a trajectory spatiotemporal overlap map.

[0035] S502: Call the trajectory spatiotemporal overlap map, extract the center position of the detection area, the area boundary and the target identifier in the common frame segment of each trajectory combination, filter the trajectory combinations whose center position falls into the boundary of the other's area and whose target identifiers are different, and group the trajectory identifier, common frame segment, overlapping area, movement direction, gray scale change direction and water area index into the same data unit to obtain the duplicate target pointing cluster. The trajectory spatiotemporal overlap map index link directory layer is invoked to perform unpacking and perspective analysis on any paired trajectory anomaly combination that is locked and bound due to meeting collision conditions. The shared time slot frame set period experienced by this locked, suspicious group is extracted, and the two-dimensional position coordinates of the absolute center of individual spatial activity, the boundary array sequence of the diffusion restriction range, and the distinguishing identification information code are extracted within the boundary of this period. A dual-track overlap and cross-checking procedure is initiated. On the test section of frame 160, the center geometric coordinates of two different identifiers are located and obtained. The coordinate array of the outer protective perimeter restriction wall of one side is extracted as a verification mold. The center core point under the other side's name is subjected to point domain inclusion relationship geometric verification and comparison. The verification conclusively confirms that the core point of one side is successfully captured and falls into the outer boundary of the other side; conversely, it verifies that the position of the other core point also falls within the outer domain controlled by the other side. Subsequently, the attribute values ​​of the identifier name codes of these two sets of information are further compared and screened in reverse, determining that their respective name strings are not of the same category. The two parties whose center points fall within each other's frame boundaries and who carry different target identifier codes are merged to form a joint abnormal trajectory dual body. The code set, shared experience interval table, overlapping boundary coordinate set, translational direction vector, grayscale status marker code column, and unique numbered water area marker are all merged into a dedicated data containment unit for centralized storage and containment, resulting in a duplicate target pointing cluster.

[0036] S503: Based on the repetitive target pointing cluster, call the trajectory spatiotemporal overlap map, compare the gray-scale change direction, movement direction and common frame segment position change relationship of each trajectory in the same data unit, retain the trajectory identifiers that are continuous before and after the common frame segment and have records of start and end frames, intra-frame detection area, movement direction, gray-scale change direction and water area index, merge the detection area and frame order of the remaining trajectories into the retained identifiers, delete the duplicate target identifiers, and obtain the cross-domain target recognition deduplication result.

[0037] The system utilizes memory channels to load and decode the subset of duplicate target-pointing cluster data packets encapsulated and incorporated into the containment group from the previous process. It then uses a horizontal lookup table to retrieve the underlying stub information of the trajectory spatiotemporal overlap map. This allows for comparison and identification between the core actions and environmental interaction indicators of each suspected multiple redundant trajectory object under review. The comparison reveals that within the same involved shared track overlapping frame sequence interval, the grayscale variation patterns of their respective positions and postures consistently point to the same marker symbols. Furthermore, within the shared span period, the degree of unidirectional motion obtained from their acquisitions scores 0.99 in the cosine verification system, indicating a highly convergent movement direction. The changes in position, distance, pace, and displacement are opposite without misalignment. Based on this, it is determined that both are entirely due to spatial line-of-sight overlap, resulting in two sets of records representing the same entity. The process of retaining, canceling, and cleaning up duplicate entries was initiated. The duration and continuity of records across different numbered registers were analyzed. It was found that one register covered a unique preceding operational period with all elements complete and rigorous, while another redundant register lacked a complete, long-chain archive of isolated periods. Therefore, it was determined that the unique account identifier with the longest and most complete record should be retained. A mandatory merging process was initiated. The remaining un-merged segments of the redundant entries that had been withdrawn and merged were transferred and migrated to the selected baseline retention identifier's account holder's affiliation register as child nodes for continuation. Redundant twin identifiers with empty shells after stripping the substantive affiliation files were deleted and cleaned up, resulting in cross-domain target identification and deduplication.

[0038] A system for cross-domain target recognition and deduplication in UAV imagery, the system comprising: The grayscale migration module acquires video frames of water patrol captured by the UAV's onboard camera, extracts the pixel grayscale information of the Yangtze finless porpoise detection area, classifies the pixel distribution according to grayscale level, extracts the direction of grayscale distribution change in continuous video frames, and associates the continuous area of ​​grayscale change direction with the detection area to obtain cross-domain grayscale change features. The trajectory extension module extracts the center position change information of the Yangtze finless porpoise detection area in continuous video frames based on cross-domain gray-scale change features, extracts the target movement direction based on the position change direction of adjacent video frames, extends the change path of the detection area along the movement direction, identifies the intermittent disappearance state of the target in the video frame, and obtains the target movement extension trajectory by combining the position change relationship before and after disappearance with the gray-scale change direction. The target retrieval module extracts the target's detection position and direction of movement before disappearance based on the target's motion trajectory, deduces the target's reappearance position along the direction of movement, forms a search area centered on the predicted position, expands the search area based on the change in the duration of the target's disappearance, extracts the center position of the Yangtze finless porpoise detection area in subsequent video frames, compares the spatial relationship between the center position of the detection area and the predicted position, filters the position change-related areas, and completes target matching by combining the direction of grayscale change to obtain the candidate related state of the disappeared target. The trajectory merging module extracts target trajectory segments from the drone cruise video based on the candidate association state of the disappeared target, filters associated segments according to the temporal connection and spatial continuity relationship between trajectory segments, connects continuous trajectory positions along the movement direction, filters associated trajectory segments by combining the movement direction change and grayscale change direction, merges the target identifiers corresponding to multiple trajectory segments, and obtains cross-frame continuous target trajectories. The deduplication module extracts target recognition information from cross-water patrol images based on continuous target trajectories across frames. It filters duplicate target trajectories based on the temporal and spatial relationships of different trajectories, extracts duplicate pointing targets by combining grayscale change direction, removes duplicate target information, and obtains cross-domain target recognition deduplication results.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for cross-domain target recognition and deduplication in UAV imagery, characterized in that, Includes the following steps: S1: Acquire video frames from the drone, extract grayscale information of the Yangtze finless porpoise detection area, classify pixels according to grayscale levels, extract grayscale distribution changes in continuous video frames, identify continuous areas of grayscale change direction and associate them with the detection area to obtain cross-domain grayscale change features. S2: Based on the cross-domain grayscale change features, extract the change in the center position of the detection area in continuous video frames, identify the target movement direction between adjacent frames, extend the position change path along the movement direction, and combine the target disappearance to form an interruption state and the grayscale change direction to form a motion continuation path, thereby obtaining the target motion extension trajectory. S3: Based on the target motion extension trajectory, identify the target's detection position and movement direction before disappearance, deduce the reappearance position along the movement direction, form a spatial search area centered on the predicted position, gradually expand the search area according to the target's disappearance duration, extract newly appearing detection areas in subsequent video frames, combine spatial positional relationships and grayscale change direction to filter candidate targets, complete target matching, and obtain the candidate association state of the disappeared target. S4: Based on the candidate correlation state of the disappeared target, extract the target trajectory segments in the cruise video, identify the temporal connection relationship and spatial continuity relationship of the trajectory segments, extend the trajectory position along the movement direction, and filter the associated segments by combining the consistency of the movement direction and the consistency of gray scale changes, and merge them into the same target identifier to obtain the cross-frame continuous target trajectory.

2. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that: The cross-domain grayscale change features include grayscale level distribution, feature change amplitude, and temporal correlation attributes. The target motion extension trajectory includes trajectory node sequence, motion range boundary, and trajectory direction features. The disappearing target candidate correlation state includes candidate target attributes, correlation credibility, and target correspondence. The cross-frame continuous target trajectory includes trajectory number, target identity attribute, and trajectory continuity marker.

3. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that: The steps for obtaining the cross-domain grayscale change features are as follows: S101: Acquire video frames of the Yangtze River surface captured by the UAV's airborne camera, extract the pixel coordinates and pixel grayscale information corresponding to the Yangtze finless porpoise detection area in each video frame, classify the pixels in the detection area according to the grayscale level, aggregate the pixel positions corresponding to the same grayscale level into a grayscale pixel set, arrange each grayscale pixel set according to the video frame acquisition order, and associate the grayscale level with the detection area identifier to obtain a grayscale pixel distribution matrix. S102: Based on the grayscale pixel distribution matrix, extract the set of pixels corresponding to the same grayscale level according to the video frame arrangement order, compare the distribution range and distribution state of each grayscale level in adjacent video frames, assign change direction markers to the expansion, contraction and stable states respectively, verify whether the direction markers remain in the same direction in consecutive video frames, and arrange the same-direction change markers, corresponding grayscale levels and frame order in association to obtain the grayscale distribution change sequence. S103: Call the grayscale distribution change sequence and the grayscale pixel distribution matrix to map the continuous direction markers to the corresponding video frame pixel positions. Aggregate regions with consistent grayscale change directions based on the spatial connection relationship of adjacent pixels. Filter regions with connected positions and the same direction markers in consecutive video frames. Spatially correspond each region to the Yangtze finless porpoise detection region in the same frame. Collect the detection region identifier, grayscale level, change direction and frame sequence information to obtain cross-domain grayscale change features.

4. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that: The steps for obtaining the target motion extension trajectory are as follows: S201: Based on the cross-domain grayscale change features, extract the boundary coordinates of the Yangtze finless porpoise detection area in continuous video frames. Use the median of the horizontal coordinates of the left and right boundaries as the center horizontal coordinate, and the median of the vertical coordinates of the upper and lower boundaries as the center vertical coordinate. Arrange the center coordinates according to the frame order, compare the signs of the center coordinate differences of adjacent frames in the horizontal and vertical directions of the image, and collect the corresponding direction signs, frame order, region identifiers and center coordinates to obtain the center displacement vector set. S202: Call the center displacement vector set and the cross-domain grayscale change feature, search the detection area corresponding to the same target identifier along the frame sequence, identify the time period when the target identifier does not appear in adjacent frames and reappears in subsequent frames, extract the end center coordinates, end movement direction, grayscale change direction and missing frame segments before disappearance, compare the spatial orientation of the center coordinates of the reproduced area relative to the end movement direction, merge the same orientation information with the missing frame segments, and obtain the target interruption association table; S203: According to the target interruption association table, call the center displacement vector set, continue the detection area change path along the movement direction of the end before disappearance, arrange the missing frame segment path nodes according to the change trend of adjacent center coordinates before disappearance, filter the target that is first in the spatial distance sorted with the end of the continued path and has the same gray change direction, and splice the path before disappearance, the missing frame segment path nodes and the reconstructed path according to the frame order, and use the same target identifier to obtain the target motion extension trajectory.

5. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that: The steps for obtaining the candidate association state of the disappeared target are as follows: S301: Based on the target motion extension trajectory, extract the end detection position before disappearance, the end movement direction and the adjacent trajectory nodes of the end, compare the relationship between the displacement direction of the adjacent trajectory nodes and the end movement direction, translate the position change between nodes with the same direction frame by frame along the end movement direction to the subsequent frames, arrange the predicted positions according to the frame order, and collect the target identifier, interruption start frame, end detection position, movement direction and the predicted positions of each frame to obtain the reconstructed position prediction domain; S302: Call the reproducible position prediction domain, expand the search boundary around each inferred position frame by frame according to the number of frames the target identifier interrupts, extract the detection area that first enters the corresponding search boundary in subsequent video frames and does not use the original target identifier, filter the first region of the spatial interval between the center position of the detection area in each frame and the inferred position in the same frame, and collect the region identifier, center position, appearance frame order and search boundary to obtain the reproducible target candidate set; S303: Based on the proposed set of reproducible target candidates, the cross-domain grayscale change features and the reproducible position prediction domain are called to compare the grayscale change direction of each candidate region with the grayscale change direction of the target before disappearance. Candidate regions with the same direction are associated with the original target identifier. The target identifier, prediction position, reproducible position, interrupted frame segment, candidate region identifier and grayscale change direction are merged in frame order to obtain the disappearing target candidate association state.

6. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that: The steps for obtaining the continuous target trajectory across frames are as follows: S401: Based on the candidate association state of the disappeared target, extract the start frame, end frame, first position, last position, movement direction and grayscale change direction of each target trajectory segment in the cruise video. Arrange the trajectory segments according to the start frame. Compare the frame order of the end frame of the previous segment with the start frame of the next segment. Mark the cases of adjacent frame order and interruption intervals respectively. Extend the last position of the previous segment frame by frame according to the interrupted frame segment and correspond it with the first position of the next segment to obtain the trajectory connection relationship table. S402: Call the trajectory connection relationship table and the candidate association state of the disappearing target, extract the end movement direction, the beginning movement direction, the position extension direction and the gray change direction of each connection segment, filter the trajectory segments with the same movement direction, the beginning position of the next segment is in the extension direction of the previous segment and the gray change direction is consistent, and classify the previous segment identifier, the next segment identifier, the connection frame segment and the direction corresponding state into the same association unit to obtain the continuous segment association chain; S403: Based on the continuous segment association chain, call the trajectory connection relationship table, arrange the trajectory segments in the same association chain according to the segment start frame and end frame, connect the end position of the previous segment, the interrupted frame segment and the beginning position of the next segment in sequence, filter the association chain where the movement direction and grayscale change direction of adjacent segments are consistent, replace the target identifier of the first segment of the starting frame with the identifier of the other segments, collect the position and direction information of each frame, and obtain the cross-frame continuous target trajectory.

7. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 1 is characterized in that, The method further includes step S5: S5: Extract target recognition information from the UAV cruise image based on the cross-frame continuous target trajectory, identify the temporal overlap and spatial coincidence between different trajectories, filter duplicate targets by combining grayscale change consistency, remove duplicate target trajectories, and obtain cross-domain target recognition deduplication results; The cross-domain target recognition and deduplication results include target differentiation information, duplicate target set, and valid recognition records.

8. The method for cross-domain migration target recognition and deduplication of UAV imagery according to claim 7 is characterized in that: The steps for obtaining the cross-domain target recognition deduplication result are as follows: S501: Extract target identification information from the UAV's cross-water cruise image based on the cross-frame continuous target trajectory, and collect the trajectory start and end frames, intra-frame detection areas, movement direction, grayscale change direction and water index according to the target identifier. Compare the frame sequence intersection of different trajectories with the overlap of detection areas in the same frame, and arrange the trajectories that have a common frame sequence and whose detection areas overlap to obtain a trajectory spatiotemporal overlap map. S502: Call the aforementioned trajectory spatiotemporal overlap map, extract the center position of the detection area, the area boundary and the target identifier within the common frame segment of each trajectory combination, filter the trajectory combinations whose center position falls into the boundary of the other's area and whose target identifiers are different, and group the trajectory identifier, common frame segment, overlapping area, movement direction, grayscale change direction and water area index into the same data unit to obtain the duplicate target pointing cluster; S503: Based on the repeated target pointing cluster, the trajectory spatiotemporal overlap map is called to compare the gray-scale change direction, movement direction and common frame segment position change relationship of each trajectory in the same data unit. The trajectory identifiers that are continuous before and after the common frame segment and have records of start and end frames, intra-frame detection area, movement direction, gray-scale change direction and water area index are retained. The detection areas and frame order of the remaining trajectories are merged into the retained identifiers. The repeated target identifiers are deleted to obtain the cross-domain target recognition deduplication result.

9. A system for cross-domain target recognition and deduplication in UAV imagery, characterized in that, The system is used in the UAV image cross-domain migration target recognition and deduplication method according to any one of claims 1-8. The system includes... The grayscale migration module acquires video frames of water patrol captured by the UAV's onboard camera, extracts the pixel grayscale information of the Yangtze finless porpoise detection area, classifies the pixel distribution according to grayscale level, extracts the direction of grayscale distribution change in continuous video frames, and associates the continuous area of ​​grayscale change direction with the detection area to obtain cross-domain grayscale change features. The trajectory extension module extracts the center position change information of the finless porpoise detection area in continuous video frames based on the cross-domain grayscale change features, extracts the target movement direction based on the position change direction of adjacent video frames, extends the change path of the detection area along the movement direction, identifies the intermittent disappearance state of the target in the video frame, and obtains the target movement extension trajectory by combining the position change relationship before and after disappearance with the grayscale change direction. The target retrieval module extracts the target's detection position and movement direction before disappearance based on the target's motion extension trajectory, deduces the target's reappearance position along the movement direction, forms a search area centered on the predicted position, expands the search area range based on the change in the target's disappearance duration, extracts the center position of the Yangtze finless porpoise detection area in subsequent video frames, compares the spatial relationship between the center position of the detection area and the predicted position, filters the position change-related areas, and completes target matching by combining the grayscale change direction to obtain the candidate related state of the disappeared target. The trajectory merging module extracts target trajectory segments from the UAV cruise video based on the candidate association states of the disappeared target, filters associated segments according to the temporal connection and spatial continuity relationship between trajectory segments, connects continuous trajectory positions along the movement direction, filters associated trajectory segments by combining the movement direction change and grayscale change direction, merges the target identifiers corresponding to multiple trajectory segments, and obtains cross-frame continuous target trajectories. The deduplication identification module extracts target identification information from the cross-waterway cruise image based on the cross-frame continuous target trajectory, filters duplicate target trajectories according to the temporal and spatial relationships of different trajectories, extracts duplicate pointing targets by combining the grayscale change direction, removes duplicate target information, and obtains the cross-domain target identification deduplication result.

Images