Vision-based detection and tracking method, system and medium for small aerial targets of unmanned aerial vehicles

By combining global detectors and local trackers on the embedded platform, using Kalman filters and lightweight models for detection and tracking of small-target drones, the problem of poor detection results on low-computing power platforms is solved, and the problem of poor detection results on low-computing power platforms is achieved. Efficient and fast drone detection and tracking.

CN118657810BActive Publication Date: 2024-11-22GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411147205.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22
Estimated Expiration
2044-08-21

AI Technical Summary

Technical Problem

The prior art is difficult to effectively detect and track small-target drones on embedded platforms with low computing power and low memory, and there are problems such as large computing complexity and large model parameters, and it is difficult to deploy, especially in the case of occlusion. good.

Method used

Using a combination of global detector and local tracker, adaptive search and prediction are performed through Kalman filters, combined with the lightweight YOLOv5s model and LightTrack framework, reduce the model parameter quantity and calculation complexity, and achieve the small-target drone Quick search and accurate tracking.

Benefits of technology

It realizes efficient detection and tracking of small-target drones on the resource-constrained embedded platform, improves detection accuracy and speed, can realize real-time processing on low-computing equipment, and supports processing speeds of more than 30fps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118657810B_ABST
    Figure CN118657810B_ABST
Patent Text Reader

Abstract

The present invention discloses a method, system and medium for detecting and tracking small targets in the air of a drone based on vision, including: obtaining a real-time detection video stream from the perspective of the drone, constructing a global detector, performing visual detection and motion detection in a key image frame, and locating the area where the small target drone is located in the key image frame as an area of ​​interest through global search; performing an adaptive search area based on the area of ​​interest, and estimating the position of the small target drone in the next key image frame through a Kalman filter; constructing a local tracker to track the search area predicted by the adaptive search of the Kalman filter, and judging the tracking result, and tracking the small target drone according to the judgment result. The global detector and the local tracking fusion algorithm of the present invention realize the detection and tracking of small target drones, improve the detection accuracy and rate of small target drones, and meet the real-time requirements.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude safety control and small-target drone detection and tracking, and more specifically, to a vision-based drone aerial small-target detection and tracking method, system and medium. Background Art

[0002] In the existing technologies, due to the complex background in the real environment, the small sample size of small target drones, and the large sampling rate of most models, it is difficult to learn the feature information of small targets, which affects the detection and tracking effect of small targets. In addition, most of the detection and tracking algorithms in the existing technologies have problems such as large model parameters and high computational complexity, which means that most models and algorithms can only run on PC platforms with higher requirements and are difficult to deploy and run on embedded platforms. Furthermore, although some of the existing technologies can be deployed on embedded platforms through model conversion, they are extremely dependent on CPU performance and cannot meet the requirements of detection and tracking accuracy and speed in actual working environments. And when there is occlusion, the algorithm cannot accurately detect and track the target. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a vision-based method, system and medium for detecting and tracking small targets in the air of unmanned aerial vehicles, which solves the problem that small target unmanned aerial vehicles are difficult to detect under low computing power and low memory embedded platforms. The detection and tracking accuracy and rate of small targets are improved by fusing the global detector and the local tracker, and by reducing the number of model parameters and computational complexity, it can be deployed under resource-constrained embedded Al platforms, thereby better meeting the efficient detection and tracking of unmanned aerial vehicles in the process of security prevention and control perception and confrontation.

[0004] The first aspect of the present invention provides a vision-based method for detecting and tracking small targets in the air of a drone, comprising:

[0005] Acquire a real-time detection video stream from the perspective of the drone, extract key image frames from the real-time detection video stream, and preprocess the key image frames;

[0006] Constructing a global detector, using the global detector to perform visual detection and motion detection in the key image frame, and locating the area where the small target drone is located in the key image frame as the area of ​​interest through global search;

[0007] An adaptive search area is performed based on the region of interest, and a position of the small target UAV in the next key image frame is estimated by a Kalman filter;

[0008] A local tracker is constructed to track the search area predicted by the Kalman filter adaptive search, and a judgment is made on the tracking result, and the small target UAV is tracked according to the judgment result.

[0009] In this solution, key image frames of the real-time detection video stream are extracted and preprocessed, specifically:

[0010] Obtain a real-time detection video stream in a current aerial scene captured by a monocular camera of a drone, calculate optical flow features of each image frame in the real-time detection video stream, obtain optical flow energy of a moving area based on the optical flow features, and weight the optical flow energy according to a moving direction of the moving area;

[0011] By comparing and analyzing the weighted optical flow energy with the preset optical flow energy threshold, key image frames in the real-time detection video stream are screened;

[0012] The key image frame is preprocessed based on a gated context aggregation network, fog features are extracted for defogging, and a defogging image of the key image frame is output.

[0013] In this solution, a global detector is constructed, and visual detection and motion detection are performed using the global detector in the key image frame, specifically:

[0014] A global detector is constructed using an appearance detection module and a motion detection module. The appearance detection module is constructed by improving Yolov5 by adding an attention mechanism and a small target detection layer.

[0015] The key image frame is used as the model input, and the channel attention and spatial attention are used to perform feature weighting in the backbone network. The weighted feature map is fused and imported into the detection head, and a small target detection layer is added to the detection head and combined with the detection layers of the other three scales for score detection;

[0016] If the score is not lower than the preset score threshold, the output of the appearance detection module is used for detection and tracking. If the score is lower than the preset score threshold, the motion detection module is used for detection.

[0017] In this solution, the motion detection module is specifically:

[0018] The sparse optical flow method is used to supplement the real-time detection video stream of the camera, and the motion area is screened out by the inter-frame difference method. The motion area is imported into the drone classifier for score detection. The drone classifier is trained with a low-speed and small drone dataset.

[0019] The area where the UAV classifier screening score is greater than the preset score threshold is taken as the area of ​​interest of the small target UAV for subsequent detection.

[0020] In this solution, an adaptive search area is performed based on the region of interest, specifically:

[0021] The Kalman filter is used to track the relative speed of the small target UAV and predict the estimated speed of the small target UAV, which is expressed as:

[0022] ;

[0023] in They represent the speed of the center of the small target UAV in the X-axis and Y-axis directions and the acceleration of the center of the small target UAV in the X-axis and Y-axis directions respectively, M represents the state transfer matrix, represents the modeling error, represents the predicted measurement value, H represents the measurement matrix, represents the noise matrix;

[0024] The position of the drone in the current key image frame is calculated by using the estimated speed, and the position in the next frame is predicted based on the estimated speed and the position, a new search area is set according to the position in the next frame, and the new search area is sent to the local tracker.

[0025] In this solution, the local tracker is specifically:

[0026] The local tracker is composed of the LightTrack framework and the judgment module. The search area obtained by the Kalman filter adaptive search is input into the LightTrack twin network. The tracking area is initialized in the search area, and the similarity score between the tracking area and the tracking window is judged.

[0027] If the similarity score is greater than the preset similarity score threshold, it is directly output; if the similarity score is in the preset similarity interval, it is judged by the drone classifier and the small target drone is tracked; if neither is met, it is directly returned to the global detector.

[0028] The second aspect of the present invention also provides a vision-based UAV aerial small target detection and tracking system, comprising: a memory, a processor, a data preprocessing module, a global detector module, an adaptive search module and a local tracker module, wherein the memory and the processor store and execute a vision-based UAV aerial small target detection and tracking method program;

[0029] The data preprocessing module extracts key image frames of the real-time detection video stream, preprocesses the key image frames, extracts fog features for defogging, and outputs defogging images of the key image frames;

[0030] The global detector module uses the global detector to perform visual detection and motion detection, and locates the area where the small target drone is located in the key image frame as the area of ​​interest through global search;

[0031] The adaptive search module is based on the adaptive search area of ​​the Kalman filter, by predicting the local area where the small target drone appears in the next frame in the global image;

[0032] The local tracker module tracks the search area predicted by the adaptive search module and makes a judgment on the tracking result to achieve tracking of the small target UAV.

[0033] The third aspect of the present invention also provides a computer-readable storage medium, which includes a vision-based UAV aerial small target detection and tracking method program. When the vision-based UAV aerial small target detection and tracking method program is executed by a processor, the steps of the vision-based UAV aerial small target detection and tracking method are implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) Through the design of the global detector and local tracker modules, the present invention can better realize the rapid search of the global area where the small target UAV is located and the rapid tracking of the local area where the small target UAV is located, reduce the amount of model calculation, and thus improve the accuracy of small target recognition.

[0036] (2) Since the appearance detection module in the global detector of the present invention adopts a lightweight YOLOv5s model for training, and the local tracker adopts a lightweight tracking algorithm model for tracking, the overall framework requires less computation and parameters, and can be quantitatively deployed on an embedded Al chip with lower computing power, thereby quickly realizing the practical application of small target UAV detection and tracking.

[0037] (3) Due to the use of a lightweight model and a global-local detection and tracking architecture, the present invention greatly reduces the number of network layers, parameters, and calculations, thereby improving the computing efficiency. In real-time embedded boards, a real-time processing speed of more than 30 fps can be achieved through the acceleration of the NPU or GPU. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0039] Figure 1 A flowchart of a vision-based method for detecting and tracking small targets in the air of a UAV is shown;

[0040] Figure 2The flowchart of the vision-based UAV aerial small target detection and tracking method is shown;

[0041] Figure 3 A flowchart of global search for UAV targets using a global detector is shown;

[0042] Figure 4 A flow chart showing local tracking of a UAV target by a local tracker is shown;

[0043] Figure 5 The diagram shows the architecture of a vision-based UAV aerial small target detection and tracking system. DETAILED DESCRIPTION

[0044] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0046] Figure 1 , 2 The flowchart and process framework diagram of the vision-based UAV aerial small target detection and tracking method are shown respectively.

[0047] like Figure 1 As shown, this embodiment provides a vision-based method for detecting and tracking small targets in the air of a drone, including:

[0048] S102, obtaining a real-time detection video stream from the perspective of the drone, extracting key image frames from the real-time detection video stream, and preprocessing the key image frames;

[0049] S104, constructing a global detector, performing visual detection and motion detection in the key image frame using the global detector, and locating the area where the small target drone is located in the key image frame as the area of ​​interest through global search;

[0050] S106, performing an adaptive search area based on the region of interest, estimating the position of the small target UAV in the next key image frame through a Kalman filter;

[0051] S108, constructing a local tracker to track the search area predicted by the Kalman filter adaptive search, and making a judgment on the tracking result, and tracking the small target UAV according to the judgment result.

[0052] It should be noted that a real-time detection video stream in the current aerial scene collected by a monocular camera of a drone is obtained, the optical flow features of each image frame in the real-time detection video stream are calculated, the optical flow energy of the moving area is obtained based on the optical flow features, and the Lucas-Kanade pyramid algorithm with high real-time performance and accuracy is preferably selected to calculate the optical flow of the image frames of the real-time detection video stream, and the optical flow energy is weighted according to the moving direction of the moving area, and the moving direction is the average value of the angle between the moving target and the horizontal direction; the weighted optical flow energy is compared and analyzed with the preset optical flow energy threshold to screen the key image frames in the real-time detection video stream; the key image frames are preprocessed based on the gated context aggregation network, the fog features are extracted for defogging, and the defogging images of the key image frames are output. The defogging algorithm is used to eliminate image noise as much as possible, improve image contrast, and use peak signal-to-noise ratio and structural similarity as evaluation indicators to evaluate the defogging effect of the key image frames.

[0053] Figure 3 A flowchart of global search for UAV targets using a global detector is shown.

[0054] According to an embodiment of the present invention, a global search for drone targets is performed using a global detector, specifically:

[0055] S302, constructing a global detector using an appearance detection module and a motion detection module, wherein the appearance detection module is constructed by adding an attention mechanism and an improved Yolov5 with a small target detection layer;

[0056] S304, using the key image frame as the model input, performing feature weighting in the backbone network using channel attention and spatial attention, fusing the weighted feature map and importing it into the detection head, adding a small target detection layer in the detection head and combining it with the detection layers of the remaining three scales to perform score detection;

[0057] S306, if the score is not lower than the preset score threshold, detection and tracking are performed using the output of the appearance detection module; if the score is lower than the preset score threshold, detection is performed using the motion detection module;

[0058] S308, using a sparse optical flow method to supplement the real-time detection video stream of the camera, screening out the motion area through an inter-frame difference method, and importing the motion area into a drone classifier for score detection, wherein the drone classifier is trained using a low-speed and small drone data set;

[0059] S310, the area where the score of the drone classifier is greater than the preset score threshold is selected as the area of ​​interest of the small target drone for subsequent detection.

[0060] It should be noted that the global detector can efficiently and accurately filter out the area where the drone is located through the appearance detection module and the motion detection module. The appearance detection module uses the lightweight yolov5s model in YOLOv5 as a pre-trained model, which can be quickly quantized and deployed to the embedded board. At the same time, the appearance detection module introduces the channel attention mechanism and the spatial attention mechanism in the backbone network of the original YOLOv5 model to improve the detection performance, and adds a 160×160 scale small target detection layer in the prediction part. This layer can be combined with the other three scale detection layers to improve the detection ability of small targets at a long distance. Load the preprocessed data set and pre-trained model, use the Adam optimizer for training, update the model parameters and save the model. The motion detector is based on the inter-frame difference method. The sparse optical flow method is used to supplement the camera, and then the inter-frame difference method is used to filter out the moving area, and then the drone classifier is used to filter the moving area to be the score of the drone. The low-speed, slow, small drone dataset used to train the drone classifier takes into account a variety of weather conditions, such as sunny, foggy, cloudy, and rainy days, as well as a variety of background scenes such as trees and near-ground conditions, enabling the classifier to perform well in a variety of weather conditions and in detections involving complex backgrounds.

[0061] In a preferred embodiment, the input image frame is first detected by the appearance detection module. If the score is higher than 0.6, the output is sent to the local tracker; if the score is lower than 0.6, it is detected by the motion detection module. The motion detection module first screens out the area of ​​moving objects, and then the drone classifier screens out the area classified as a drone with a score greater than 0.6 and sends it to the local detection module. If neither of the above is successful, the next frame detection is performed.

[0062] It should be noted that the adaptive search area estimates the position of the target in the next frame through the Kalman filter, and uses the Kalman filter to track the relative speed of the small target drone to predict the estimated speed of the small target drone, which is expressed as:

[0063] ;

[0064] in They represent the speed of the center of the small target UAV in the X-axis and Y-axis directions and the acceleration of the center of the small target UAV in the X-axis and Y-axis directions respectively, M represents the state transfer matrix, represents the modeling error, represents the predicted measurement value, H represents the measurement matrix, represents the noise matrix;

[0065] Because the position change is highly nonlinear when the camera moves quickly, the relative speed is continuous in this case, and the Kalman filter is used to track the relative speed of the drone target. The Kalman filter estimates the target state through a linear difference equation. The Kalman filter algorithm is mainly divided into two stages: prediction and measurement update. In the update step, if the detection is successful, the target state is updated with the actual measurement value; if the detection fails, the target state is updated with the best prediction value of the previous frame. The position of the drone in the current key image frame is calculated by obtaining the estimated speed, and the position in the next frame is predicted based on the estimated speed and the position. A new 300×300 search area is set with the predicted target position as the center in the next frame, and the new search area is sent to the local tracker. By predicting the local area where the target can appear in the next frame in the global image, the calculation amount of the model is reduced, so that the model can track the target faster.

[0066] Figure 4 A flow chart of local tracking of UAV targets by a local tracker is shown.

[0067] According to an embodiment of the present invention, local tracking of a drone target is performed by a local tracker, specifically:

[0068] S402, using the LightTrack framework and the judgment module to form a local tracker, input the search area obtained by the Kalman filter adaptive search into the LightTrack twin network, initialize the tracking area in the search area, and judge the similarity score between the tracking area and the tracking window;

[0069] S404, if the similarity score is greater than the preset similarity score threshold, it is directly output; if the similarity score is in the preset similarity interval, it is judged by the drone classifier and the small target drone is tracked; if neither is met, it is directly returned to the global detector.

[0070] It should be noted that the local tracker is composed of a lightweight tracking module LightTrack and a judgment module. By inputting the search area obtained by the Kalman filter adaptive search into the twin network of LightTrack, the similarity judgment between the initialized tracking area and the tracking window is performed. In order to avoid the occlusion or large-scale feature transformation problems that occur when the drone is in flight during the tracking process, the similarity score is judged. If the similarity score is greater than 0.2, it is directly output; if the similarity score is between 0.1 and 0.2, it is judged by the drone classifier; if none of them meet the requirements, it is directly returned to the global detector.

[0071] LightTrack uses Neural Architecture Search (NAS) to design a lighter and more efficient target tracker, which can achieve superior performance compared to hand-designed SOTA trackers, while requiring much less computation and parameters, and can run faster when deployed on resource-constrained mobile chips. All possible network architectures are encoded into only one backbone supernet and one head supernet, where the backbone supernet is pre-trained on ImageNet and then fine-tuned with tracking data, while the head supernet is trained directly with tracking data. The supernet only needs to be trained once, and then each candidate structure inherits its weights directly from the supernet, with the training accuracy and model complexity as supervision guidance, and the architecture search is performed on the trained supernet. A pair of tracking images is taken as input, including an example image and a search image. The example image represents the object of interest, while the search image represents the search area in the subsequent video frame. Both inputs are processed by the pre-trained backbone network for feature extraction. The two feature maps generated are cross-correlated to generate correlation volumes. The head supernet contains a classification branch and a regression branch for target localization.

[0072] Figure 5 The diagram shows the architecture of a vision-based UAV aerial small target detection and tracking system.

[0073] This embodiment provides a vision-based UAV aerial small target detection and tracking system 5, comprising: a memory 51, a processor 52, a data preprocessing module 53, a global detector module 54, an adaptive search module 55 and a local tracker module 56, wherein the memory and the processor store and execute a vision-based UAV aerial small target detection and tracking method program;

[0074] The data preprocessing module extracts key image frames of the real-time detection video stream, preprocesses the key image frames, extracts fog features for defogging, and outputs defogging images of the key image frames;

[0075] The global detector module uses the global detector to perform visual detection and motion detection, and locates the area where the small target drone is located in the key image frame as the area of ​​interest through global search;

[0076] The adaptive search module is based on the adaptive search area of ​​the Kalman filter, by predicting the local area where the small target drone appears in the next frame in the global image;

[0077] The local tracker module tracks the search area predicted by the adaptive search module and makes a judgment on the tracking result to achieve tracking of the small target UAV.

[0078] The present embodiment provides a computer-readable storage medium, which includes a vision-based UAV aerial small target detection and tracking method program. When the vision-based UAV aerial small target detection and tracking method program is executed by a processor, the steps of the vision-based UAV aerial small target detection and tracking method are implemented.

[0079] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0080] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0081] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0082] Those skilled in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0083] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0084] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A vision-based method for detecting and tracking small targets in the air of a UAV, characterized in that: The following steps are involved: Acquire a real-time detection video stream from the perspective of the drone, extract key image frames from the real-time detection video stream, and preprocess the key image frames; Constructing a global detector, using the global detector to perform visual detection and motion detection in the key image frame, and locating the area where the small target drone is located in the key image frame as the area of ​​interest through global search; An adaptive search area is performed based on the region of interest, and a position of the small target UAV in the next key image frame is estimated by a Kalman filter; Construct a local tracker to track the search area predicted by the Kalman filter adaptive search, make a judgment on the tracking result, and track the small target UAV according to the judgment result; Construct a global detector, and use the global detector to perform visual detection and motion detection in the key image frame, specifically: A global detector is constructed using an appearance detection module and a motion detection module. The appearance detection module is constructed by improving Yolov5 by adding an attention mechanism and a small target detection layer. The key image frame is used as the model input, and the channel attention and spatial attention are used to perform feature weighting in the backbone network. The weighted feature map is fused and imported into the detection head, and a small target detection layer is added to the detection head and combined with the detection layers of the other three scales for score detection; If the score is not lower than the preset score threshold, the output of the appearance detection module is used for detection and tracking. If the score is lower than the preset score threshold, the motion detection module is used for detection; The motion detection module is specifically: The sparse optical flow method is used to supplement the real-time detection video stream of the camera, and the motion area is screened out by the inter-frame difference method. The motion area is imported into the drone classifier for score detection. The drone classifier is trained with a low-speed and small drone data set. The area with a score greater than a preset score threshold screened by the drone classifier is taken as the area of ​​interest of the small target drone for subsequent detection.

2. According to the vision-based UAV aerial small target detection and tracking method of claim 1, it is characterized in that: Extract key image frames of the real-time detection video stream and preprocess the key image frames, specifically: Obtain a real-time detection video stream in a current aerial scene captured by a monocular camera of a drone, calculate optical flow features of each image frame in the real-time detection video stream, obtain optical flow energy of a moving area based on the optical flow features, and weight the optical flow energy according to a moving direction of the moving area; By comparing and analyzing the weighted optical flow energy with the preset optical flow energy threshold, key image frames in the real-time detection video stream are screened; The key image frame is preprocessed based on a gated context aggregation network, fog features are extracted for defogging, and a defogging image of the key image frame is output.

3. The method for detecting and tracking small targets in the air of a UAV based on vision according to claim 1 is characterized in that: Based on the region of interest, an adaptive search region is performed, specifically: The Kalman filter is used to track the relative speed of the small target UAV and predict the estimated speed of the small target UAV, which is expressed as: , , in , Respectively represent the center of the small target drone Axis and The speed in the axis direction and the center of the small target drone Axis and The acceleration in the axial direction, represents the state transfer matrix, represents the modeling error, represents the predicted measured value, represents the measurement matrix, represents the noise matrix; The position of the drone in the current key image frame is calculated by using the estimated speed, and the position in the next frame is predicted based on the estimated speed and the position, a new search area is set according to the position in the next frame, and the new search area is sent to the local tracker.

4. The method for detecting and tracking small targets in the air of a UAV based on vision according to claim 1 is characterized in that: The local tracker is specifically: The local tracker is composed of the LightTrack framework and the judgment module. The search area obtained by the Kalman filter adaptive search is input into the LightTrack twin network. The tracking area is initialized in the search area, and the similarity score between the tracking area and the tracking window is judged. If the similarity score is greater than the preset similarity score threshold, it is directly output; if the similarity score is in the preset similarity interval, it is judged by the drone classifier and the small target drone is tracked; if neither is met, it is directly returned to the global detector.

5. A vision-based UAV aerial small target detection and tracking system, characterized in that: Implementing the vision-based UAV aerial small target detection and tracking method as described in any one of claims 1 to 4, comprising: a memory, a processor, a data preprocessing module, a global detector module, an adaptive search module and a local tracker module, wherein the memory and the processor store and execute a vision-based UAV aerial small target detection and tracking method program; The data preprocessing module extracts key image frames of the real-time detection video stream, preprocesses the key image frames, extracts fog features for defogging, and outputs defogging images of the key image frames; The global detector module uses the global detector to perform visual detection and motion detection, and locates the area where the small target drone is located in the key image frame as the area of ​​interest through global search; The adaptive search module is based on the adaptive search area of ​​the Kalman filter, by predicting the local area where the small target drone appears in the next frame in the global image; The local tracker module tracks the search area predicted by the adaptive search module and makes a judgment on the tracking result to achieve tracking of the small target UAV.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a vision-based UAV aerial small target detection and tracking method program. When the vision-based UAV aerial small target detection and tracking method program is executed by a processor, the steps of the vision-based UAV aerial small target detection and tracking method as described in any one of claims 1 to 4 are implemented.