A scale adaptive target tracking method in unmanned airport scene

Abstract

The application discloses a scale adaptive target tracking method in an unmanned aerial vehicle scene, and is used for solving the problems of tracking frame drift and scale estimation inaccuracy caused by the change of visual angle, target deformation and occlusion in the aerial photography of the unmanned aerial vehicle. The core steps include: obtaining a first frame of a video and establishing an initial template of a target; determining a search area of a current frame based on the position of the target in a previous frame, extracting features and calculating a response map to obtain a preliminary center position of the target; innovatively taking the preliminary position as the center to construct a two-dimensional candidate frame set containing different scale factors and different length-width ratio factor combinations; performing feature matching and response value calculation on each candidate frame one by one, and selecting the optimal one as the final position and boundary frame of the target in the current frame to realize the adaptive joint estimation of the scale and the ratio; and the application significantly improves the precision and robustness of target tracking in the unmanned aerial vehicle scene through the scale-ratio joint adaptive estimation and four-dimensional confidence double-branch processing mechanism.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, UAV visual perception and moving target tracking technology, and specifically relates to a scale-adaptive target tracking method in UAV scenarios. Background Technology

[0002] With the rapid development and widespread application of drone technology, visual target tracking based on drone platforms has demonstrated tremendous application value in fields such as aerial reconnaissance, intelligent security, traffic monitoring, environmental monitoring, and agricultural plant protection. However, the inherent characteristics of drone aerial photography scenarios bring a series of severe challenges to target tracking technology: First, changes in drone flight altitude and camera zoom can cause rapid and continuous changes in target size; second, the relative motion between the drone and the target can lead to significant changes in the observation perspective; third, the background under the drone's perspective is often complex, with interference from similar objects; in addition, targets are easily obscured for short periods by obstacles such as buildings and trees, or may move out of the search range due to rapid movement, resulting in tracking box drift or even tracking failure.

[0003] Currently, existing target tracking methods can be mainly divided into two categories: those based on correlation filtering and those based on deep learning.

[0004] Tracking methods based on correlation filtering (such as KCF, DSST, and SAMF) improve tracking speed by performing cyclic matrix operations in the frequency domain. To address scale variation issues, the DSST algorithm introduces a one-dimensional scale filter. By sampling the target's central region at different scales, it trains and updates an independent scale filter, achieving a proportionally scaled estimate of the target. The SAMF algorithm utilizes a scale pooling strategy, selecting the largest response from multiple proportionally scaled candidate boxes as the optimal scale. The main problem with these methods is that they can only handle proportionally scaled targets and cannot cope with non-proportional scale deformations where the target's aspect ratio changes independently. In UAV overhead or oblique views, the shape of targets (such as vehicles and pedestrians) often exhibits independent changes in width and height due to the changing viewpoint. The rectangular tracking boxes of the aforementioned methods often fail to closely fit the target, easily introducing background interference or losing some target information, leading to a decrease in tracking accuracy.

[0005] In recent years, deep learning-based target tracking methods (such as SiamRPN, DiMP, and OSTrack) have made significant progress. For example, Chinese patent publication CN119887855B proposes a UAV target tracking method that uses noise-adaptive Kalman filtering to predict target position and state. It proposes a confidence-based feature extraction and judgment strategy, performing IoU matching between the trajectory and detection box. For targets with a matching degree below a threshold and a confidence change rate above a threshold, it uses the lightweight feature re-identification network Rep-OSNet to extract appearance features; otherwise, it reuses features from the previous frame. It designs a cost function that fuses motion direction and appearance similarity to perform cascaded matching between the target and the confirmed trajectory. It uses a detector to obtain the target bounding box and confidence level, and achieves multi-target trajectory association through IoU matching and cascaded matching, focusing on data association and identity maintenance in multi-target scenarios. For example, Henan University of Science and Technology published a Chinese patent in 2025, CN121053164A, on a UAV target tracking method based on variable template tokens and predictive relocalization. This method uses three indicators—confidence score, peak-to-sidelobe ratio, and size change rate—to evaluate tracking quality and proposes a predictive relocalization strategy based on Kalman filtering and a Transformer model to address target occlusion. While this method shows improvement in handling occlusion, its scale estimation strategy still mainly relies on proportional scaling or implicit Transformer learning, without explicitly constructing an adaptive estimation model for independent changes in aspect ratio. Therefore, the tracking box fit still has room for improvement when facing complex changes in target shape in UAV scenarios. Furthermore, this type of method uses specific deep learning architectures (such as Transformer) for relocalization, requiring high hardware computing power and limiting its real-time performance and versatility on airborne embedded platforms.

[0006] In summary, existing technologies lack effective explicit solutions to the problem of simultaneous continuous scale changes and independent aspect ratio changes of targets in drone scenarios. Furthermore, there is still room for optimization in the dimensions and template update strategies of their confidence assessment systems, making it difficult to achieve highly reliable and persistent tracking in complex occlusion scenarios. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and propose a scale-adaptive target tracking method and device in UAV scenarios. By establishing a scale-proportion joint adaptive candidate box evaluation mechanism, a multi-dimensional confidence decision system, and a linked dual-branch template update and motion prediction relocation strategy, the invention achieves accurate adaptive estimation of target scale and aspect ratio, and reliably recovers the target when tracking quality deteriorates due to occlusion, thereby significantly improving the accuracy and robustness of target tracking in UAV scenarios.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: A scale-adaptive target tracking method in a drone scenario includes the following steps: S1, acquire the UAV video image sequence, initialize the bounding box of the target to be tracked in the first frame image, extract the target region features, and establish the initial target template; S2, Based on the target position and target motion state of the previous frame, determine the target search area in the current frame image and extract features from the search area; S3, match the features of the current frame search region with the target template to obtain the target response map, and determine the preliminary center position of the target in the current frame based on the peak value of the response map; S4. Using the initial center position as the center, construct a candidate box set that is jointly adaptive in terms of scale and proportion; the candidate box set contains multiple candidate bounding boxes with different scale factors and different combinations of aspect ratio factors. S5, extract the image region features corresponding to each candidate bounding box, match them with the target template, calculate the response value corresponding to each candidate bounding box, and select the candidate bounding box with the largest response value as the final position and scale of the target in the current frame, so as to achieve adaptive estimation of scale and proportion. S6, calculate a comprehensive confidence level of tracking reliability by combining four dimensions: peak response, response stability, scale change rate, and target motion continuity of the current frame; S7, when the overall confidence level is higher than a preset threshold, the target template and target motion state are updated using the first update strategy; when the overall confidence level is lower than or equal to the preset threshold, it is determined that the target is occluded or the tracking is drifting, a motion prediction model is started to predict the possible position of the target in the current frame, and the target is searched again in a set area around the predicted position. At the same time, the target template is updated or the update is paused using the second update strategy, and the update weight of the second update strategy is lower than that of the first update strategy. S8, repeat S2 to S7 until all frames are tracked.

[0009] Furthermore, the candidate bounding boxes mentioned in step S4 specifically include one or more combinations of the following: original scale candidate boxes, proportionally enlarged candidate boxes, proportionally reduced candidate boxes, candidate boxes with only width changes, and candidate boxes with only height changes.

[0010] Furthermore, the comprehensive confidence level in step S6 is calculated by weighting and summing the normalized response peak value, the response stability characterizing the fluctuation of the response map, the scale change rate characterizing the change in the target box area between adjacent frames, and the motion continuity characterizing the deviation between the actual displacement and the predicted displacement of the target.

[0011] Furthermore, when the overall confidence level of tracking reliability in step S6 is lower than or equal to a preset threshold, it is determined that the target is temporarily completely occluded or out of the field of view; the motion prediction model in step S7 predicts the possible position of the target in the current frame based on the target's historical motion state and using a general motion prediction method such as Kalman filtering or constant velocity model.

[0012] Furthermore, in step S7, the first update strategy is a linear interpolation smooth update based on the tracking result of the current frame; the second update strategy is to pause the target template update or to update using a template update rate much lower than the first update strategy.

[0013] Furthermore, the target template mentioned in steps S3 and S5 is based on multi-feature fusion, and the features include at least two of the following: histogram of oriented gradients, color features, and background-aware features.

[0014] A second aspect of the present invention provides a scale-adaptive target tracking device in a drone scenario, comprising: a video acquisition module for acquiring a sequence of drone video images; The target initialization module is used to initialize the bounding box of the target to be tracked in the first frame image and extract the target region features to establish the target initial template; The preliminary localization module is used to determine the search area of ​​the current frame based on the target position of the previous frame, match the features of the search area with the target template to obtain a response map, and determine the preliminary center position of the target based on the peak value of the response map. The scale-scale joint estimation module is used to generate a set of candidate bounding boxes containing different scale factors and different aspect ratio factors, centered on the initial center position, and calculate the response value of each candidate box through feature matching, and select the best to determine the final position and bounding box of the target in the current frame. The tracking status evaluation module is used to calculate the overall confidence level of tracking reliability by comprehensively considering the peak response, response stability, scale change rate, and target motion continuity of the current frame. An adaptive processing module is used to execute a first strategy or a second strategy based on the comparison result of the comprehensive confidence level and a preset threshold. The first strategy includes updating the target template and motion state normally when the confidence level is high. The second strategy includes starting a motion prediction model to predict the target position and performing a local re-search when the confidence level is low, while reducing the template update weight or pausing the update.

[0015] Furthermore, candidate boxes with multiple scales and aspect ratios are generated from the same center point, and the aspect ratios of the candidate boxes are different from each other.

[0016] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0017] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0018] The beneficial effects of this invention are: (1) Higher tracking accuracy: joint adaptive estimation of scale and proportion: This invention introduces a joint candidate box evaluation mechanism using scale factors and aspect ratio factors after initial target localization. Compared to scale estimation methods that only perform proportional scaling, this mechanism can better adapt to situations where the target's width and height change asynchronously. By constructing a set of two-dimensional candidate boxes containing different combinations of scale factors and aspect ratio factors, and performing feature matching and response value calculation for each box, this invention can accurately capture independent changes in the target's width and height. This mechanism allows the tracking box to more closely fit the actual target contour, effectively reducing the mixing of background information and the loss of target information when the UAV's perspective changes drastically or the target undergoes non-rigid deformation, thus significantly improving tracking accuracy. This mechanism constitutes a key technical feature that distinguishes this invention from conventional proportional scale search methods.

[0019] (2) Stronger anti-occlusion and anti-drift capabilities: Four-dimensional comprehensive confidence judgment and dual-branch processing: This invention goes beyond the limitations of existing technologies that rely solely on peak response or a few indicators to judge tracking quality. By innovatively integrating four dimensions—"peak response," "response stability," "scale change rate," and "motion continuity"—it constructs a more comprehensive and robust integrated confidence assessment system for tracking reliability. This system can more delicately and accurately perceive the critical states of target occlusion or tracking drift. When a decision is made as unreliable, it immediately triggers a predictive relocalization branch based on a general motion model, pausing or reducing the template update weight, effectively preventing erroneous templates from "contaminating" the tracker. This ensures that when the target reappears after a short period of occlusion, the tracker can quickly and accurately recapture the target, achieving an intelligent closed loop of "memory" and "retrieval."

[0020] (3) The algorithm has good versatility and real-time performance: it does not depend on a specific complex architecture. In step S7 of this invention, motion prediction and relocalization preferably employs general prediction methods such as Kalman filtering or constant velocity models, rather than being limited to a specific deep learning architecture (such as Transformer). This makes the method of this invention not only accurate but also simple in structure, low in computational overhead, and more versatile. It is particularly suitable for deployment and real-time operation on UAV embedded platforms with limited computing power, and has good prospects for engineering applications.

[0021] (4) High long-term template purity: Hierarchical template update strategy: By strictly linking the template update method to the confidence level, and implementing a differentiated strategy of "normal update for high scores and pause or slow update for low scores", the target template library is always filled with high-quality and high-purity target feature information. This avoids "noise" features introduced by short-term occlusion, motion blur or mislocalization, laying a solid foundation for long-term stable tracking. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the overall process of a scale-adaptive target tracking method in an unmanned aerial vehicle (UAV) scenario according to the present invention.

[0024] Figure 2 This is a schematic diagram of the target scale-proportion joint adaptive estimation of the present invention.

[0025] Figure 3 This is a schematic diagram of the occlusion processing and relocation process based on confidence level judgment in this invention.

[0026] Figure 4 This is a schematic diagram illustrating motion prediction and relocation under target occlusion conditions according to the present invention.

[0027] Figure 5 This is a schematic diagram of the target feature extraction and reliable template update structure of the present invention. Detailed Implementation

[0028] 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, and 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.

[0029] Example 1: Stable tracking of clearly visible targets This embodiment describes the complete process of accurate adaptive tracking using the method of the present invention when the target remains visible and undergoes significant changes in scale and proportion. Please refer to... Figure 1 (Overall flowchart), the specific steps of this embodiment are as follows: S1: First Frame Initialization and Template Establishment The drone's high-definition camera captures a sequence of aerial video images. The operator or an automated detection algorithm provides an initial rectangular bounding box for a target to be tracked (e.g., a moving white car) in the first frame of the video. ,in With the center coordinates, and These are the width and height of the frame, respectively.

[0030] The system extracts features from the image region within the bounding box. Please refer to [link / reference]. Figure 5 (Feature Extraction Structure Diagram) This embodiment employs a multi-feature fusion strategy, specifically extracting Histogram of Oriented Gradients (HOG) and Color Histogram (CN) features. HOG features excel at describing the shape and edge contours of a target and are insensitive to changes in illumination; color features, on the other hand, capture the target's color distribution information. These two features are fused to form a robust target appearance model, which serves as the initial target template. .

[0031] S2: Search Region Determination and Feature Extraction After tracking to the number After the frame, the known target location is and motion state (such as the velocity of the target in the image plane) For the current frame (the first frame) The system uses the target position from the previous frame as the center and crops a rectangular region in the image that is K times larger than the target bounding box (e.g., K=3) as the search area. This is done to ensure that even if the target is displaced, it remains within the analysis range.

[0032] Then, for the image patch within the search area, dense feature extraction is performed using the same HOG and CN feature extraction methods as in S1 to obtain the feature map of the search area in the current frame. .

[0033] S3: Response Map Calculation and Preliminary Center Location The feature map of the search region in the current frame With the target initial template Perform a cross-correlation operation. This operation can be efficiently performed in the frequency domain using a Fast Fourier Transform, generating a response map that describes the similarity between each location within the search region and the target.

[0034] The location of the peak in the response map represents the center of the image region in the current frame that is most similar to the initial template. Therefore, by mapping this peak location back to the original image coordinates, the preliminary center location of the target in the current frame can be determined. At this point, the positioning is a rough positioning based on the original target scale, without taking into account changes in scale and proportion.

[0035] S4: Construction of Scale-Proportion Joint Adaptive Candidate Box Set This step is the core element that distinguishes this invention from existing technologies. Please refer to [link / reference]. Figure 2 (Schematic diagram of scale-proportion joint estimation), after obtaining the preliminary center position Instead of simply using a few proportionally varying boxes to evaluate the optimal scale, the system constructs a two-dimensional candidate box set. This set encompasses combinations of different scale factors s and different aspect ratio factors r.

[0036] Specifically, this embodiment defines the following candidate box generation rules: Scale factor s: Set S = {0.9, 1.0, 1.1}, representing the overall scaling level relative to the target size in the previous frame.

[0037] Dimension factor r: Set R = {0.8, 1.0, 1.2}, representing the adjustment factor for the ratio of the target box width to its height (w / h).

[0038] By combining the Cartesian product of s and r, and keeping the candidate box area basically unchanged or scaling it according to s, a total of 9 candidate boxes can be generated. For example: Candidate box 1 (s=1.0, r=1.0): the original scale and scale box.

[0039] Candidate box 2 (s=1.0, r=1.2): The area remains basically unchanged, but the width increases and the height decreases. It is suitable for situations where the target changes from the side to the front.

[0040] Candidate box 3 (s=1.1, r=0.8): The overall size is enlarged, with increased height and decreased width, which is suitable for situations where the target is approaching and viewed from a top-down perspective.

[0041] And so on.

[0042] The center point of all candidate boxes is aligned with the initial center position. .like Figure 2 As shown, these candidate boxes have various shapes, including... Figure 2 The examples of "scaling up proportionally" and "scaling down proportionally" also include boxes for non-proportional changes such as "increasing width, increasing height, decreasing width, and decreasing height," thus providing complete coverage of the various forms that the car target may take in this embodiment.

[0043] S5: Candidate Box Selection and Final Positioning For each candidate bounding box constructed in S4, the system performs the following operations: 1. Based on the coordinates of the candidate boxes, map them back to the image of the current frame.

[0044] 2. Crop the image of the specified area and scale it to match the target template. Same size.

[0045] 3. Extract the HOG and CN fusion features of the image in this region.

[0046] 4. Compare this feature with the target template Perform a dot product operation to calculate a response value. This value represents the degree of similarity between the content of the candidate box area and the target template.

[0047] Iterate through all 9 candidate boxes and compare their response values. The candidate box with the highest response value is selected as the final tracking result for the target in the current frame. .Should It not only has the optimal center position, but also the optimal width and height to accurately match the current shape of the target. For example, when the target car turns, its body length in the image becomes shorter and its width becomes larger. The candidate box with a scaling factor r=1.2 will receive the highest response and will be accurately selected.

[0048] S6: Four-Dimensional Comprehensive Confidence Assessment Obtain the optimal tracking box for the current frame Then, the system begins to evaluate the reliability of this tracking and calculates a comprehensive confidence level. Please see. Figure 3 (Confidence level assessment flowchart) This embodiment extracts indicators from the following four dimensions: 1. Peak response : That is, the response value of the winning candidate box in S5 The result is normalized and used as an indicator. The higher the peak value, the better the match.

[0049] 2. Response stability The fluctuation of the response value in the local region centered on the peak in the response map is usually measured by the peak-side lobe ratio (PSR). The higher the PSR, the sharper the peak, the more reliable the tracking result, and the less susceptible it is to interference from similar objects in the surrounding area.

[0050] 3. Rate of change at scale : Calculate the values ​​of two adjacent frames (the first frame) Frame and the The ratio of the area of ​​the target bounding box to the area of ​​the target bounding box (frame) is compared with the historical average rate of scale change to measure the smoothness of scale change. Abrupt scale changes are often a signal of tracking errors.

[0051] 4. Motion continuity Based on the historical trajectory of the previous few frames, a Kalman filter is used to predict the target's position in the current frame, and the Euclidean distance between the actual observed position and the predicted position in the current frame is calculated. The smaller the deviation, the better the motion continuity, and the more reliable the tracking results.

[0052] Overall confidence level The result is obtained by weighted summation of these four indicators: (in, to As the preset weights, due to and This is a deviation index, which is negatively correlated with the confidence level, so it is taken as a negative value.

[0053] S7: Confidence-based two-branch decision processing The calculated overall confidence level With preset experience threshold (For example Compare the values ​​(=0.6) and execute the two-branch strategy: Situation A: > (Reliable tracking) like Figure 3 As indicated by the "Yes" branch, the system determines that the tracking result of the current frame is accurate and reliable.

[0054] Template Update: Execute the first update strategy for the target template. Perform a smooth update. For example... Figure 5 As shown in the left branch, linear interpolation is used: .in, These are the fused features extracted from the winning candidate boxes in the current frame. For template update rate (e.g.) =0.02), this strategy ensures that the template can slowly adapt to the slow appearance changes of the target.

[0055] Motion state update: Update the target position and velocity information in the current frame as input for prediction models such as Kalman filtering in the next frame.

[0056] Situation B: ≤ (Tracking may be unreliable) like Figure 3 The "No" branch indicates that the system determines that the target may be occluded, out of the field of view, or experiencing tracking drift.

[0057] Motion Prediction and Relocalization: The relocalization mechanism is initiated immediately. The system ignores the localization results of the response map in the current frame and instead uses a general motion prediction model based on historical motion states (in this embodiment, a Kalman filter) to predict the most likely location of the target in the current frame. Then, with Centered on the target, a smaller local search region is generated. Within this region, the search and scale-scale joint estimation steps S3 to S5 are repeated in order to re-locate the target. (See also...) Figure 4 (Diagram of occlusion prediction and relocation) Figure 4 The goal was shown in the , The frame is visible, and it is occluded in the t-th frame (at which point the confidence level is...). (It may drop sharply), the system activates the motion prediction model, and uses historical trajectories (in the direction of the arrows in the figure) to predict its potential location in frame t; in the second frame... Once the target reappears in the frame, tracking is resumed through a local re-search.

[0058] Template Update: Execute the second update strategy. To prevent occlusions or backgrounds from being incorrectly learned into the target model, pause template updates at this point (i.e., let...). =0), or use a very small update rate (e.g. =0.002) to perform a "conservative" update, thereby maintaining the purity of the target template.

[0059] S8: Iterative Loop Determine if the current frame is the last frame of the video sequence. If not, return to step S2, and based on the final tracking result of the previous frame, continue processing the next frame until continuous and stable tracking of the entire video sequence is completed.

[0060] Example 2: Tracking based on a richer set of candidate boxes and three-feature fusion This embodiment, based on Embodiment 1, further expands the richness and feature types of the candidate box set to cope with more complex and varied drone aerial photography scenarios.

[0061] S1 and S2: Basically the same as in Example 1. The difference is that, in order to enrich the feature representation, this example uses a three-feature fusion method of HOG features, color features (CN), and background-aware features in the feature extraction part when establishing the initial template in S1. Background-aware features can learn the differences between the target and its surrounding background, and can more significantly distinguish the target from complex backgrounds.

[0062] S3: Same as in Example 1, calculate the response map and obtain the preliminary center position C.

[0063] S4: Construct a more refined set of candidate boxes Please refer to it again. Figure 2 The candidate boxes in Example 1 are for illustrative purposes only. In this example, to achieve precise capture of the target shape, a denser set of two-dimensional candidate boxes is defined: The set of scale factors s is expanded to: S = {0.85, 0.92, 1.0, 1.08, 1.15}, with a total of 5 levels.

[0064] The set of aspect ratio factors r is expanded to: R = {0.75, 0.87, 1.0, 1.13, 1.25}, with a total of 5 levels.

[0065] By combining these two sets, a total of 5 × 5 = 25 candidate bounding boxes with different shapes are generated. These candidate boxes not only vary in overall size, but more importantly, their aspect ratios also range from "short and wide" to "tall and narrow," enabling more precise matching of the appearance proportion changes of targets (such as pedestrians and cyclists) in running, jumping, turning, and other states. The center point of all 25 candidate boxes is aligned with the initial center position C.

[0066] S5: Candidate box optimization based on three-feature fusion The 25 candidate boxes generated by S4 are cropped and scaled to the template size one by one, and then three features (HOG + CN + background-aware features) are extracted. The fused features of each candidate box are matched with the target template that also contains the three features, and the comprehensive response value is calculated. Finally, the candidate box with the highest response value is selected as the most representative target state of the current frame.

[0067] Compared to the nine candidate boxes in Example 1, this process involves more computation, but its adaptability to target deformation is also enhanced.

[0068] S7: Model Selection in Occlusion Handling In this embodiment, the four-dimensional comprehensive confidence assessment process of S6 remains unchanged.

[0069] When the system enters the "tracking may be unreliable" branch of S7, its motion prediction model preferentially selects the constant velocity model. This model assumes that the target's motion is uniform linear over a short period of time, is computationally simple and efficient, and has extremely low computational requirements.

[0070] Based on the target's position in the previous two frames, the system calculates the target's average velocity vector V in the image coordinate system. Predicted position. .

[0071] When the target is in a subsequent frame (e.g.) Figure 4 The first in When a frame reappears and is recaptured by the local re-search mechanism, the system detects it. Restore to That's all. At this point, the system will not immediately update the template with the full update rate α, but will first update it with a reduced update rate (such as...). =0.008) Update several frames, waiting for several consecutive frames (e.g., 3 frames) All are above the stricter threshold. (For example Only after reaching 0.75 did it return to the standard update rate. =0.02. This "slow recovery" mechanism effectively prevents the template from being impacted by slight positioning deviations during recapture.

[0072] Example 3: Rapid response to re-acquiring a target after it has completely left the field of view This embodiment focuses on illustrating the processing capability of the method of the present invention when the target completely leaves the current field of view and then reappears. This is an extension of the occlusion handling in Embodiments 1 and 2 to address extreme cases.

[0073] Scene Description: A drone is tracking a robot moving on the ground. At an intersection, the robot is completely obscured by a large building and disappears from the drone's view (out of the field of view) for more than ten frames. The robot then reappears from another direction of the building.

[0074] S1-S5: During the robot's visibility period, the system accurately tracks it according to the method described in Example 1, and updates its motion state and template in real time.

[0075] S6: A sudden drop in confidence triggers an out-of-field decision. When the robot completely disappears behind the building, the response value in S5 becomes extremely low regardless of which candidate box is selected. In the confidence assessment of S6, the peak response value... and response stability It will drop sharply. At the same time, the deviation between the position predicted by the Kalman filter and the optimal (but incorrect) response position found will increase dramatically, leading to motion continuity issues. The indicators surged.

[0076] The combined effect of multiple dimensions of indicators increases the overall confidence level. Rapidly falling below the preset threshold Unlike the temporary occlusion in Examples 1 and 2, if For multiple consecutive frames (e.g., more than 5 frames), the value remains below a lower sub-threshold. (For example If the value is 0.2, the system will make a judgment that "the target has left the field of view".

[0077] S7: Prediction and Global Search Preparation in the Field of View Motion prediction: After a "out-of-view" determination, the system stops performing any appearance-based searches in the current frame (because the target is no longer in the frame, forcing a search would only increase false alarms). The system relies entirely on the motion prediction model. In practical applications, if the UAV's own position and attitude information is known, the motion prediction model will also integrate the UAV's pose change data to more accurately predict the target's position in the "world coordinate system," then map it back to the image plane to estimate the image area where it might reappear.

[0078] Template freeze: Immediately implement the strictest second update strategy—completely suspend template updates. =0), completely "freeze" the target template in the high-quality state of the last frame before it leaves the field of view.

[0079] Target reappearance and recapture Please refer to it again. Figure 4 It can be compared to the first The frame is out of view, the first... The frame is the target frame that reappears.

[0080] In the In the next frame, the robot re-enters the field of view from the other end of the building. At this point, the system is in predictive search mode after "out of view". Instead of relying on the full-image response map, the system sets up a local search window around the image region where the robot is most likely to appear, as estimated by the motion prediction model.

[0081] Within this local window, the system utilizes a frozen, clean template to perform precise searches and scale-proportion joint estimations from S3 to S5. Because the template is uncontaminated, the matching response value spikes instantly once the robot appears within the prediction region. Rapidly rebounded to The tracker is immediately reactivated and accurately locks onto the target's new location and scale. Afterward, the system switches back to the "Reliable Tracking" branch and begins updating the template at a normal or gradually recovering update rate.

[0082] This embodiment fully demonstrates the robustness of the invention from "stable tracking" to "occlusion / out-of-view judgment", then to "motion prediction and memory retention", and finally to "local precise recapture", which is especially suitable for recovery tracking scenarios after the target has disappeared for a long time.

[0083] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A scale-adaptive target tracking method in a UAV scenario, characterized in that, Includes the following steps: S1, acquire the UAV video image sequence, initialize the bounding box of the target to be tracked in the first frame image, extract the target region features, and establish the initial target template; S2, Based on the target position and target motion state of the previous frame, determine the target search area in the current frame image and extract features from the search area; S3, match the features of the current frame search region with the target template to obtain the target response map, and determine the preliminary center position of the target in the current frame based on the peak value of the response map; S4, let the width of the target box in the previous frame be... Height is The scaling factor is The aspect ratio factor is The candidate box width and height are respectively , Thus, while maintaining the scale factor to control the overall area change, the width and height are adaptively adjusted in reverse through the aspect ratio factor; S5, extract the image region features corresponding to each candidate bounding box, match them with the target template, calculate the response value corresponding to each candidate bounding box, and select the candidate bounding box with the largest response value as the final position and scale of the target in the current frame, so as to achieve adaptive estimation of scale and proportion. S6, calculate a comprehensive confidence level of tracking reliability by combining four dimensions: peak response, response stability, scale change rate, and target motion continuity of the current frame; S7, when the overall confidence level is higher than a preset threshold, the target template and target motion state are updated using the first update strategy; When the overall confidence level is lower than or equal to a preset threshold, it is determined that the target is occluded or the tracking is drifting. A motion prediction model is started to predict the possible position of the target in the current frame, and the target is searched again in a set area around the predicted position. At the same time, the target template is updated or paused with a second update strategy. The update weight of the second update strategy is lower than that of the first update strategy. S8, repeat S2 to S7 until all frames are tracked.

2. The scale-adaptive target tracking method in a UAV scenario according to claim 1, characterized in that, In step S4, the candidate bounding box set includes at least candidate boxes with proportional scale changes and candidate boxes with non-proportional length and width changes; wherein, the candidate boxes with non-proportional length and width changes include at least two types of candidate boxes with increasing width and decreasing height, candidate boxes with decreasing width and increasing height, candidate boxes with only width changes, and candidate boxes with only height changes.

3. The scale-adaptive target tracking method in a UAV scenario according to claim 1, characterized in that, The comprehensive confidence level mentioned in step S6 The calculation method is as follows: ,in, For the normalized peak response, In response to stability, The scale reliability index is derived from the scale variation deviation between adjacent frames. The motion continuity index is derived from the deviation between the actual and predicted positions of the target. .

4. The scale-adaptive target tracking method in a UAV scenario according to claim 1, characterized in that, When the overall confidence level of tracking reliability in step S6 is lower than or equal to a preset threshold, it is determined that the target is temporarily completely occluded or out of the field of view; the motion prediction model in step S7 is based on the target's historical motion state and uses a general motion prediction method such as Kalman filtering or constant velocity model to predict the target's possible position in the current frame.

5. The scale-adaptive target tracking method in a UAV scenario according to claim 1, characterized in that, In step S7, the first update strategy is a linear interpolation smooth update based on the tracking result of the current frame; the second update strategy is to pause the target template update or to update using a template update rate much lower than the first update strategy.

6. The scale-adaptive target tracking method in a UAV scenario according to claim 1, characterized in that, The target template mentioned in steps S3 and S5 is based on multi-feature fusion, and the features include at least two of the following: histogram of oriented gradients, color features, and background-aware features.

7. A scale-adaptive target tracking device for unmanned aerial vehicle (UAV) scenarios, characterized in that, include: The video acquisition module is used to acquire video image sequences from the drone. The target initialization module is used to initialize the bounding box of the target to be tracked in the first frame image and extract the target region features to establish the target initial template; The preliminary localization module is used to determine the search area of ​​the current frame based on the target position of the previous frame, match the features of the search area with the target template to obtain a response map, and determine the preliminary center position of the target based on the peak value of the response map. The scale-scale joint estimation module is used to generate a set of candidate bounding boxes containing different scale factors and different aspect ratio factors, centered on the initial center position, and calculate the response value of each candidate box through feature matching, and select the best to determine the final position and bounding box of the target in the current frame. The tracking status evaluation module is used to calculate the overall confidence level of tracking reliability by comprehensively considering the peak response, response stability, scale change rate, and target motion continuity of the current frame. An adaptive processing module is used to update the target template and target motion state according to a first update strategy when the overall confidence level is higher than a preset threshold; when the overall confidence level is lower than or equal to the preset threshold, the motion prediction model is activated to predict the target position, and a local re-search is performed around the predicted position, while the template update rate is reduced or the template update is paused according to a second update strategy.

8. The apparatus according to claim 7, characterized in that, The candidate bounding box set generated by the scale-scale joint estimation module includes: candidate boxes with multiple scales and multiple aspect ratios generated at the same center point, and the aspect ratio values ​​of the candidate boxes are different from each other.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 6.

Images