Target tracking method for unmanned aerial vehicle

Abstract

The invention discloses an unmanned aerial vehicle target tracking method, relates to the field of computer vision, and aims to solve the problems of target shielding, rapid movement and efficiency in a dynamic scene. Comprising the following steps: acquiring a bounding box and confidence of a target through a detector, and predicting the position and state of the target by using noise adaptive Kalman filtering; putting forward a feature extraction judgment strategy based on confidence, performing IoU matching on the trajectory and the detection frame, extracting appearance features of a target of which the matching degree is lower than a threshold value and the confidence change rate is higher than the threshold value by using a lightweight feature re-recognition network Rep-OSNet, otherwise, multiplexing the features of the previous frame; designing a cost function fusing the motion direction and appearance similarity to perform cascade matching on the target and the confirmation state trajectory; carrying out secondary association matching on the mismatched target and track in cascade matching; and updating a matched track state by using noise adaptive Kalman filtering, deleting a mismatched track which is lost for a long time, and outputting a target track prediction frame and an ID (Identity).

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and unmanned aerial vehicle application, and specifically relates to an unmanned aerial vehicle target tracking method, which is suitable for real-time multi-target tracking and trajectory management in dynamic scenes. Background Art

[0002] In key areas such as modern security and military defense, the widespread use of drones has brought many potential risks and challenges. Drone target tracking technology has become an important research direction that needs to be developed and improved.

[0003] Multi-target tracking includes two parts: multi-target detection and multi-target tracking. Classic deep learning target detection networks are divided into two-stage networks and single-stage networks. Among them, the two-stage detection algorithm has high detection accuracy but slow running speed; the single-stage detection algorithm runs fast but has a high false alarm rate. Two-stage networks, such as RCNN, FastR-CNN, FasterR-CNN and CascadeR-CNN, first need to generate candidate regions, and then classify and locate the candidate regions, which are suitable for applications with higher detection accuracy requirements; single-stage networks, such as SSD, YOLO series and CenterNet, directly generate coordinate positions and class probabilities, which are faster than two-stage networks. Multi-target tracking methods based on deep learning are mainly divided into two categories: Tracking Based Detection (TBD) and Joint Detection Tracking (JDT). The JDT algorithm attempts to integrate the detection module and the tracking module to improve the inference speed, but in practical applications, it faces the problem of difficulty in co-training of each module, resulting in unstable overall performance. The TBD strategy has a clear multi-stage design structure and can optimize detection and tracking separately, and performs well in terms of adaptability to complex scenes. The classic TBD tracker SORT proposed a simple and real-time data cascading method. DeepSORT adds appearance information based on the SORT framework to improve the algorithm performance and achieve tracking of targets that are occluded for a long time. StrongSORT further upgrades it from the aspects of detection, embedding and cascading, optimizes DeepSORT with the latest components and training techniques, and proposes an appearance-free link model (AFLink) that only uses spatiotemporal information to cascade short trajectories into complete trajectories, as well as Gaussian smoothing interpolation (GSI) to compensate for missing detections. BoT-SORT achieves better bounding box positioning through camera motion compensation and more accurate Kalman filter state vectors, as well as a new fusion method based on cosine distance of IoU and re-id. However, in the face of complex problems such as blurred background interference, occlusion and disappearance of objects, there are still defects such as inaccurate tracking and difficulty in re-locking after the target is lost.

[0004] In summary, the existing multi-target tracking technology still has many problems, especially in terms of target detection accuracy, tracking stability and adaptability to resource-constrained environments in complex environments. Therefore, it is urgent to develop a new UAV target tracking method. Summary of the invention

[0005] In view of the shortcomings and deficiencies of the above method, the present invention proposes a UAV target tracking method, the steps are:

[0006] Step S1, drone target detection: obtain the bounding box and confidence of the drone target in the current frame through the drone target detector;

[0007] Step S2, state prediction: using the noise adaptive Kalman filter algorithm to predict the position and motion state of the target in the current frame;

[0008] Step S3, feature extraction judgment: propose a feature extraction judgment strategy based on confidence, perform IoU matching on the trajectory and the detection box, and calculate the confidence change rate of the target, mark the target whose IoU matching degree is lower than the threshold and whose confidence change rate is higher than the threshold, and reuse the previous frame feature for the target whose IoU matching degree is higher than the threshold or whose confidence change rate is lower than the threshold;

[0009] Step S4, extracting appearance features: by constructing a lightweight feature re-identification network Rep-OSNet, extracting appearance features for targets whose IoU matching degree in the current frame is lower than a threshold and whose confidence change rate is higher than a threshold;

[0010] Step S5, cascade matching: cascade matching is performed on the detected target and the confirmed state trajectory by designing an adaptive cost function that integrates the motion direction and the appearance similarity;

[0011] Step S6, secondary association matching: performing secondary association matching on the target and trajectory that have cascade matching mismatch;

[0012] Step S7, state update: Use noise adaptive Kalman filtering to update the matching trajectory state, delete the trajectory that is still mismatched in the secondary association match and whose loss time is greater than the threshold; output the updated target trajectory prediction box and ID.

[0013] Further preferably, the drone target detection network in step S1 specifically includes the following steps:

[0014] S11, build drone target detection network:

[0015] Construct feature extraction backbone network;

[0016] Construct a multi-scale feature fusion network;

[0017] Build the detection head.

[0018] S12, using the NMS algorithm to process the redundant detection boxes output by the network and output the detection results.

[0019] Further preferably, in step S2, the position and motion state of the target in the current frame are predicted using a noise adaptive Kalman filter algorithm, which specifically includes the following steps:

[0020] S21, predicting the state of the target at the current moment according to the motion model of the target and the state estimation value at the previous moment.

[0021] S22, calculate the correlation metric between the predicted state and the currently detected target position, set the correlation metric threshold to 0.7, compare the correlation metric and the threshold, and determine whether the trajectory is confirmed or unconfirmed. If it is less than the threshold, it is confirmed, otherwise it is unconfirmed.

[0022] Further preferably, in step S3, a confidence-based feature extraction decision strategy is proposed, the trajectory and the detection frame are matched by IoU, and the confidence change rate of the target is calculated, and the target whose IoU matching degree is lower than the threshold and the confidence change rate is higher than the threshold is marked, and for the target whose IoU matching degree is higher than the threshold or the confidence change rate is lower than the threshold, the previous frame feature is reused; specifically, the following steps are included:

[0023] S31, calculate the IoU between the trajectory prediction box and the current frame detection box, mark the target whose IoU matching degree is lower than the threshold, and define the judgment condition as:

[0024]

[0025] in, is the predicted bounding box of the ith track, To detect bounding boxes, the threshold ; If this condition is met, the target motion state is considered stable, the appearance features of the previous frame are directly reused, the feature extraction of the current frame is skipped, and the targets with IoU matching below the threshold are marked.

[0026] S32, for targets whose IoU matching degree is lower than the threshold, further calculate their confidence change rate, which is specifically defined as follows:

[0027]

[0028] in, is the detection confidence of the current frame, is the confidence of the previous frame, threshold ;

[0029] If satisfied If the condition is not met, the target appearance is considered to have changed significantly and is marked; if not, the target appearance is considered to have not changed significantly and there is no need to re-extract features.

[0030] Further preferably, in step S4, a lightweight feature re-identification network Rep-OSNet is constructed to extract target appearance features whose IoU matching degree in the current frame is lower than a threshold and whose confidence change rate is higher than a threshold, specifically comprising the following steps:

[0031] S41, introduce the re-parameterized structure SAG into the OSNET network to construct Rep-OSNet.

[0032] S42, SAG has a multi-branch convolutional structure during the training stage.

[0033] Branch 1 is 3×3 convolution, average pooling, multi-layer perceptron MLP, and Sigmoid generates spatial attention weights;

[0034] Branch 2 is a 3×3 depth-wise separable convolution;

[0035] The output is branch 1 weight × branch 2 features + original input.

[0036] In the inference stage, the multi-branch structure is equivalently converted to a single 3×3 convolution to reduce computational complexity and resource consumption. The parameter fusion formula is:

[0037]

[0038] in, is the weight matrix of the 3×3 depthwise separable convolution in branch 2, is the weight parameter of MLP in branch 1, is the attention weight matrix after the Sigmoid activation function, which is used to spatially weight the features of branch 2.

[0039] Further preferably, in step S5, an adaptive cost function integrating motion direction and appearance similarity is designed to perform cascade matching of the detected target and the confirmed state trajectory, which specifically includes the following steps:

[0040] S51, normalize multiple matching items, calculate the IoU between the trajectory prediction box and the detection box, and define the normalized IoU distance as:

[0041]

[0042] in, is the predicted bounding box of the ith track, To detect bounding boxes;

[0043] Extract the appearance feature vector of the detected target, calculate its cosine similarity with the track history feature library, and define the appearance distance as:

[0044]

[0045] in, To detect the target appearance features, is the kth feature in the trajectory history feature library;

[0046] According to the historical motion direction of the trajectory and the relative position of the detection frame, the direction difference angle is calculated and normalized to:

[0047]

[0048] in, Predict the direction angle for the trajectory, is the azimuth of the detection box relative to the position of the previous frame on the trajectory.

[0049] S52, set the dynamic weight adaptive mechanism according to the track loss time Frame and maximum allowed loss duration , calculate the appearance feature weight ratio coefficient α:

[0050]

[0051] in, is the loss time of trajectory i; It is an adaptive loss time threshold, which is dynamically adjusted according to the scenario. is the velocity variance of trajectory i in the last 5 frames; γ=0.05 is the velocity variance adjustment factor.

[0052] Fusion of motion and appearance matching terms to generate the final matching cost function:

[0053]

[0054] in, , , and all distance items are in the range of [0,1].

[0055] S53, setting the enabling conditions for direction matching. To reduce the computational overhead, the direction matching degree Only enabled if all of the following conditions are met:

[0056] Track loss time ;

[0057] Average track speed 1.5×average target size / frame rate;

[0058] The target is not marked as stationary (historical speed variance ); In other cases, and , .

[0059] S54, perform parameter adaptive optimization, online dynamic adjustment, update every 30 frames :

[0060]

[0061] Among them, λ=1 is determined by cross-validation, using the MOT20 and VisDrone datasets, with tracking success rate (Success Rate) and frame rate (FPS) as optimization targets, and the grid search method is used to select the optimal value in the range of λ∈[0.5,1.5]. The average target size is the average width of all detection boxes in the current frame, and the average target speed is the mean of all trajectory displacement speeds.

[0062] Further preferably, in step S6, performing secondary association matching on the target and trajectory of the cascade matching mismatch specifically includes the following steps:

[0063] S61, obtaining mismatched targets and trajectories in the cascade matching stage.

[0064] S62: Use the CMC algorithm to update the position of the detection box in the association matching algorithm.

[0065] S63, set the IoU threshold T of the secondary matching to 0.5, and trigger the two-way matching verification when the IoU value between the detection box and the trajectory is greater than 0.5.

[0066] For each track, calculate the IoU value between it and all detection boxes. When the IoU value between a detection box and the track is greater than the threshold T, the detection box is considered as a candidate matching box for the track;

[0067] For each detection box, calculate the IoU value between it and all tracks. When the IoU value between a track and the detection box is greater than the threshold T, the track is considered as a candidate matching track for the detection box;

[0068] When a detection box is a candidate matching box of a certain track in the forward matching, and the track is also a candidate matching track of the detection box in the reverse matching, it is determined that the detection box matches the track successfully, otherwise the match is unsuccessful;

[0069] Further preferably, in step S7, a noise adaptive Kalman filter is used to update the matching trajectory state, the trajectories that are still mismatched and lost more than a threshold after the secondary association matching are deleted, and the updated target trajectory prediction frame and ID are output; specifically, the following steps are included:

[0070] S71, add a noise adaptation mechanism to the Kalman filter algorithm, introduce trajectory influence factors, and adjust the observation noise covariance. This process is defined as follows:

[0071]

[0072] in, is the trajectory impact factor, is the detection confidence, is the preset minimum noise covariance.

[0073] In order to control the computational complexity, this formula adopts the method of dividing the interval into The value of is dynamically smoothed, and the value is defined as follows:

[0074]

[0075] in, is the number of consecutive confirmations of the trajectory.

[0076] Through a large number of experimental verification , , , , the system can achieve the optimal balance between target tracking accuracy and computational efficiency. Specifically:

[0077] The minimum number of consecutive confirmation frames required for the initial stabilization of the corresponding target trajectory;

[0078] is the threshold value for the trajectory to be highly reliable;

[0079] and It is obtained by fitting historical data through gradient optimization algorithm.

[0080] S72, updating the status of the matching track obtained by the cascade matching and the secondary association matching.

[0081] S73, setting the trajectory loss frame number threshold to 30 frames, and deleting the trajectory when the secondary association matching trajectory loses more than 30 frames continuously.

[0082] S74, update the appearance model of each tracker to adapt to the change of the target appearance

[0083] S75, output the trajectory of each tracked target, including the target ID, bounding box and confidence.

[0084] The present invention constructs a set of efficient and accurate drone target tracking system through lightweight network design, dynamic feature extraction strategy, adaptive cost function and noise-robust filtering algorithm, which surpasses the existing technology in terms of computational efficiency, long-term tracking stability and scene adaptability, and provides reliable technical support for key fields such as security inspection and military monitoring. Compared with the existing technology, the present invention has the following beneficial effects:

[0085] 1. In view of the computational redundancy caused by the extraction of all-target features in traditional tracking algorithms, the present invention proposes a feature extraction judgment strategy based on confidence. Its characteristic is that a dual judgment mechanism is introduced. First, IoU matching is performed and then the confidence change rate is judged to screen the targets to be extracted. For targets with stable motion state or no significant change in appearance, the features of the previous frame are directly reused. This strategy not only reduces the computational overhead, but also avoids the problem of target loss caused by temporary occlusion of the target. Compared with the BOTSORT fixed extraction strategy, the present invention significantly improves the real-time performance of the system while ensuring tracking accuracy, and is particularly suitable for high-speed flight scenarios of unmanned aerial vehicles.

[0086] 2. In order to solve the problem that the traditional re-identification network has insufficient feature discrimination and high computational complexity in occluded scenes, the present invention proposes a lightweight feature re-identification network Rep-OSNet. Its characteristics are the introduction of a heavy parameterized structure SAG, the use of a multi-branch convolution structure in the training phase, and the fusion of the multi-branch structure into a single 3×3 convolution in the inference phase, which reduces the number of model parameters and improves the inference speed while maintaining the robustness of feature discrimination. Compared with the traditional OSNet network, the feature extraction efficiency of Rep-OSNet in drone scenarios is significantly optimized, solving the bottleneck problem of limited edge device resources.

[0087] 3. In order to solve the problem of long-term occlusion tracking failure caused by the fixed weight of the traditional cost function, the present invention designs a cost function that dynamically fuses motion and appearance features. Its characteristic is that the motion and appearance weights are adaptively adjusted through the trajectory loss time, velocity variance and direction consistency. The appearance weight coefficient formula is introduced. When the target loss time is prolonged or the motion uncertainty increases, the IoU matching weight is gradually reduced and the appearance feature weight is increased. At the same time, the direction matching item is only enabled when the target is high-speed and lost for a short period of time, avoiding directional noise interference of low-speed targets. This design improves the recovery success rate of the algorithm in occluded scenes, and has significant advantages over the fixed weight strategy of DeepSORT.

[0088] 4. Aiming at the problem of poor noise adaptability of Kalman filtering in complex scenes, the present invention improves its state update mechanism. Its characteristic is that it proposes a noise adaptive model, dynamically adjusts the observation noise covariance by combining the trajectory influencing factor with the detection confidence, so that even if the single-frame confidence of the target is low, the tracking continuity can still be maintained through historical performance. Compared with the NSA Kalman algorithm of StrongSORT, this method improves the trajectory integrity in blurred and semi-occluded scenes, and controls the computational complexity by interval division to avoid exponential resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0090] Figure 1 It is a structural diagram of a UAV target tracking method in an example of the present invention;

[0091] Figure 2 It is a flow chart of a method for tracking a target of an unmanned aerial vehicle in an example of the present invention;

[0092] Figure 3 It is a feature extraction and determination flow chart in an example of the present invention;

[0093] Figure 4 It is a Rep-OSNet network structure diagram in the example of the present invention;

[0094] Figure 5 is a structural diagram of a re-parameterized structure SAG in an example of the present invention; DETAILED DESCRIPTION

[0095] In order to make the objectives, technical solutions and advantages of the present invention clearer, the specific implementation process of the present invention will be described in more detail below in combination with specific implementation methods and drawings, so as to facilitate technical personnel to more accurately understand the present invention and apply it in various specific fields.

[0096] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0097] See also Figure 1 The present invention discloses a method for tracking a target of an unmanned aerial vehicle. The specific process is as follows: Figure 2 The steps are:

[0098] Step S1, obtaining the bounding box and confidence of the drone target in the current frame through the drone target detector; specifically includes the following steps:

[0099] S11, build the UAV target detection network, build the feature extraction backbone network, build the multi-scale feature fusion network and the detection head.

[0100] S12, using the collected data set, select The loss function and Adam optimizer are used to train the drone target detection network. The loss function is the weighted sum of positioning loss, classification loss and confidence loss, which is defined as follows:

[0101]

[0102] In the above formula, represents the positioning loss, represents the classification loss, Represents the confidence loss.

[0103] S13, using the NMS algorithm to process the redundant detection frame output by the network and output the detection result.

[0104] Step S2, using a noise adaptive Kalman filter algorithm to predict the position and motion state of the target in the current frame; specifically includes the following steps:

[0105] S21, predicting the state of the target at the current moment according to the motion model of the target and the estimated value of the state at the previous moment;

[0106] S22, calculate the correlation metric between the predicted state and the currently detected target position, set a threshold, compare the correlation metric and the threshold, and determine whether the trajectory is confirmed or unconfirmed. If it is less than the threshold, it is confirmed, otherwise it is unconfirmed.

[0107] Step S3, propose a feature extraction decision strategy based on confidence, such as Figure 3 As shown in the figure, the trajectory and the detection box are matched by IoU, and the confidence change rate of the target is calculated. The targets with IoU matching lower than the threshold and confidence change rate higher than the threshold are marked. For targets with IoU matching higher than the threshold or confidence change rate lower than the threshold, the features of the previous frame are reused. The specific steps include the following:

[0108] S31, calculate the IoU between the trajectory prediction box and the current frame detection box, mark the target whose IoU matching degree is lower than the threshold, and define the judgment condition as:

[0109]

[0110] in, is the predicted bounding box of the ith track, To detect the bounding box; through a large number of experiments, it is verified that when the IoU matching threshold c=0.7, it can effectively distinguish between the sudden change of the target motion state and normal jitter;

[0111] If this condition is met, the target motion state is considered stable, and the appearance features of the previous frame are directly reused, the feature extraction of the current frame is skipped, and the targets with IoU matching below the threshold are marked.

[0112] S32, for targets whose IoU matching degree is lower than the threshold, further calculate their confidence change rate, which is specifically defined as follows:

[0113]

[0114] in, is the detection confidence of the current frame, The confidence of the previous frame; the confidence change rate threshold is determined based on the statistical distribution (90% quantile) of significant changes in the target appearance ;

[0115] If satisfied If the condition is not met, the target appearance is considered to have changed significantly and is marked; if not, the target appearance is considered to have not changed significantly and there is no need to re-extract features.

[0116] Step S4, construct a lightweight feature re-identification network Rep-OSNet, such as Figure 4 As shown, the target whose IoU matching degree is lower than the threshold and whose confidence change rate is higher than the threshold marked in the above step S3 is obtained, and appearance feature extraction is performed, which specifically includes the following steps:

[0117] S41, introduce the re-parameterized structure SAG into the OSNET network to construct Rep-OSNet.

[0118] S42, SAG has a multi-branch convolutional structure in the training stage, such as Figure 5 As shown;

[0119] Branch 1 is 3×3 convolution, average pooling, multi-layer perceptron MLP, and Sigmoid generates spatial attention weights;

[0120] Branch 2 is a 3×3 depth-wise separable convolution;

[0121] The output is branch 1 weight × branch 2 features + original input.

[0122] In the inference stage, the multi-branch structure is equivalently converted to a single 3×3 convolution to reduce computational complexity and resource consumption. The parameter fusion formula is:

[0123]

[0124] in, is the weight matrix of the 3×3 depthwise separable convolution in branch 2, is the weight parameter of MLP in branch 1, is the attention weight matrix after the Sigmoid activation function, which is used to spatially weight the features of branch 2.

[0125] Step S5, cascade matching the detected target and the confirmed state trajectory by designing an adaptive cost function that integrates the motion direction and appearance similarity; specifically includes the following steps:

[0126] S51, normalize multiple matching items, calculate the IoU between the trajectory prediction box and the detection box, and define the normalized IoU distance as:

[0127]

[0128] in, is the predicted bounding box of the ith track, To detect bounding boxes;

[0129] Extract the appearance feature vector of the detected target, calculate its cosine similarity with the trajectory history feature library, and define the appearance distance as:

[0130]

[0131] in, To detect the target appearance features, is the kth feature in the trajectory history feature library;

[0132] According to the historical motion direction of the trajectory and the relative position of the detection frame, the direction difference angle is calculated and normalized to:

[0133]

[0134] in, Predict the direction angle for the trajectory, is the azimuth of the detection box relative to the position of the previous frame on the trajectory.

[0135] S52, set the dynamic weight adaptive mechanism according to the track loss time Frame and maximum allowed loss duration , calculate the appearance feature weight ratio coefficient α:

[0136]

[0137] in, is the loss time of trajectory i; It is an adaptive loss time threshold, which is dynamically adjusted according to the scenario. is the velocity variance of trajectory i in the last 5 frames; γ=0.05 is the velocity variance adjustment factor.

[0138] Fusion of motion and appearance matching terms to generate the final matching cost function:

[0139]

[0140] in, , , and all distance items are in the range of [0,1].

[0141] S53, setting the enabling conditions for direction matching. To reduce the computational overhead, the direction matching degree Only enabled if all of the following conditions are met:

[0142] Track loss time ;

[0143] Average track speed 1.5×average target size / frame rate;

[0144] The target is not marked as stationary (historical speed variance ); In other cases, and , .

[0145] in, and 1.5× target average size / frame rate, obtained by optimizing the target kinematic model.

[0146] S54, perform parameter adaptive optimization, online dynamic adjustment, update every 30 frames :

[0147]

[0148] Among them, λ=1 is determined by cross-validation, using the MOT20 and VisDrone datasets, with tracking success rate (Success Rate) and frame rate (FPS) as optimization targets, and the grid search method is used to select the optimal value in the range of λ∈[0.5,1.5]. The average target size is the average width of all detection boxes in the current frame, and the average target speed is the mean of all trajectory displacement speeds.

[0149] Step S6, performing secondary association matching on the target and trajectory of the cascade matching mismatch; specifically comprising the following steps:

[0150] S61, obtaining mismatched targets and trajectories in the cascade matching stage.

[0151] S62, using the CMC algorithm to update the position information of the detection frame in the association matching algorithm, The CMC algorithm is used for compensation, and the process is defined as follows:

[0152]

[0153] In the above formula, represents the estimated value of the target feature after compensation, Represents the original detection frame position, represents the motion compensation matrix, Represents the motion compensation term to correct the position of the detection frame.

[0154] S63, setting the IoU threshold T of the secondary matching to 0.5, triggering the two-way matching verification when the IoU value between the detection box and the trajectory is greater than 0.5;

[0155] For each track, calculate the IoU value between it and all detection boxes. When the IoU value between a detection box and the track is greater than the threshold T, the detection box is considered as a candidate matching box for the track;

[0156] For each detection box, calculate the IoU value between it and all tracks. When the IoU value between a track and the detection box is greater than the threshold T, the track is considered as a candidate matching track for the detection box;

[0157] When a detection box is a candidate matching box of a certain track in the forward matching, and the track is also a candidate matching track of the detection box in the reverse matching, it is determined that the detection box matches the track successfully, otherwise the matching is unsuccessful.

[0158] Step S7, using the noise adaptive Kalman filter to update the matching trajectory state, delete the trajectory that is still mismatched after the secondary association matching and whose loss time exceeds the threshold, and output the updated target trajectory prediction box and ID. Specifically including the following steps:

[0159] S71, add a noise adaptation mechanism to the Kalman filter algorithm, introduce trajectory influence factors, and adjust the observation noise covariance. This process is defined as follows:

[0160]

[0161] in, is the trajectory impact factor, is the detection confidence, is the preset minimum noise covariance.

[0162] In order to control the computational complexity, this formula adopts the method of dividing the interval into The value of is dynamically smoothed, and the value is defined as follows:

[0163]

[0164] in, is the number of consecutive confirmations of the trajectory.

[0165] Through a large number of experimental verification, it was found , , , , the system can achieve the optimal balance between target tracking accuracy and computational efficiency. Specifically:

[0166] The minimum number of consecutive confirmation frames required for the initial stability of the corresponding target trajectory;

[0167] is the threshold value for the trajectory to be highly reliable;

[0168] and It is obtained by fitting historical data through gradient optimization algorithm.

[0169] S72, updating the status of the matching track obtained by the cascade matching and the secondary association matching.

[0170] S73, setting the trajectory loss frame number threshold to 30 frames, and deleting the trajectory when the secondary association matching trajectory loses more than 30 frames continuously.

[0171] The setting of the trajectory loss frame number threshold of 30 frames comprehensively considers the maximum effective recall time of the target re-identification algorithm, and determines the optimal value through the ROC curve.

[0172] S74, update the appearance model of each tracker to adapt to the change of the target appearance

[0173] S75, output the trajectory of each tracked target, including the target ID, bounding box and confidence.

[0174] Obviously, the above steps of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for tracking a drone target, characterized in that: include: S1, UAV target detection: obtain the bounding box and confidence of the UAV target in the current frame through the UAV target detector; S2, state prediction: use the noise adaptive Kalman filter algorithm to predict the position and motion state of the target in the current frame; S3, feature extraction judgment: propose a feature extraction judgment strategy based on confidence, perform IoU matching on the trajectory and the detection box, and calculate the confidence change rate of the target. Mark the targets whose IoU matching is lower than the threshold and whose confidence change rate is higher than the threshold. For targets whose IoU matching is higher than the threshold or whose confidence change rate is lower than the threshold, reuse the features of the previous frame. S4, extract appearance features: by constructing a lightweight feature re-identification network Rep-OSNet, extract appearance features for targets whose IoU matching degree in the current frame is lower than the threshold and whose confidence change rate is higher than the threshold; S5, cascade matching: cascade matching is performed between the detected target and the confirmed state trajectory by designing an adaptive cost function that integrates the motion direction and appearance similarity; S6, secondary association matching: perform secondary association matching on the target and trajectory that do not match the cascade matching; S7, state update: Use the noise adaptive Kalman filter to update the matching trajectory state, delete the trajectory that is still mismatched in the secondary association match and whose loss time is greater than the threshold; output the updated target trajectory prediction box and ID.

2. The method for tracking a target by using an unmanned aerial vehicle according to claim 1, wherein: The confidence-based feature extraction decision strategy includes performing IoU matching on the trajectory and the detection box and calculating the confidence change rate. For targets whose IoU matching degree is higher than a threshold or whose confidence change rate is lower than a threshold, the feature information of the previous frame is directly used, and targets whose IoU matching degree is lower than a threshold and whose confidence change rate is higher than a threshold are marked.

3. The method for tracking a target by using an unmanned aerial vehicle according to claim 1, wherein: The lightweight feature re-identification network Rep-OSNet includes introducing a re-parameterized structure, setting a re-parameterized structure SAG in the Rep-OSNet network, having a multi-branch convolution structure in the training stage, including an input layer, an average pooling layer, an MLP, an activation function Sigmoid and an output layer, and merging into a single 3×3 convolution in the inference stage to reduce computational complexity and memory usage.

4. The method for tracking a target by using an unmanned aerial vehicle according to claim 1, wherein: In the cascade matching process, a cost function that dynamically fuses motion and appearance features is used to adaptively adjust the matching weights according to the duration of trajectory loss, consistency of motion direction, and stability of target speed. This process specifically includes: S4-1, calculation of IoU distance, appearance distance, and direction difference angle normalization: The normalized IoU distance is defined as: in, is the predicted bounding box of the ith track, To detect bounding boxes; Define the appearance distance as: in, To detect the target appearance features, is the kth feature in the trajectory history feature library; Calculate the direction difference angle and normalize it to: in, Predict the direction angle for the trajectory, is the azimuth of the detection frame relative to the position of the previous frame on the trajectory; S4-2, set up a dynamic weight adaptation mechanism, including: Calculate the appearance feature weight ratio coefficient: in, is the loss time of trajectory i; It is an adaptive loss time threshold, which is dynamically adjusted according to the scenario. is the velocity variance of trajectory i in the last 5 frames, and γ=0.05 is the velocity variance adjustment factor; S4-3, set the direction matching enable condition, only when the track is lost , average trajectory speed 1.5×average target size / frame rate and Enable ; S4-4, perform parameter adaptive optimization, update every 30 frames : in, Determine λ∈[0.8,1.2] through offline training, the average target size is the average width of all detection boxes in the current frame, and the average target speed is the mean of all trajectory displacement speeds; S4-5, Dynamic Fusion Motion and Appearance Matching, The final matching cost is defined as follows: in, , , and all distance items are in the range of [0,1].

5. The method for tracking a target by using an unmanned aerial vehicle according to claim 1, wherein: In the state update process, a noise adaptive mechanism is added to the Kalman filter algorithm, and a trajectory influencing factor is introduced to construct a noise adaptive Kalman filter, taking into account both the confidence of target detection and the historical tracking performance. This process is defined as follows: in, is the trajectory impact factor, is the detection confidence, is the preset minimum noise covariance; In order to control the computational complexity, this formula adopts a piecewise function, and the calculation method is as follows: in, is the number of consecutive confirmations of the trajectory.

Images