Lightweight anti-shielding unmanned aerial vehicle target tracking method and system

Abstract

The invention belongs to but not limited to the technical field of computer vision, and discloses a lightweight anti-shielding unmanned aerial vehicle target tracking method and system, and the method comprises the steps: obtaining a video image through an airborne camera; inputting the video into a tracking method SiamRPN + + A; obtaining a tracking result confidence map; calculating a weighted neighborhood index value for the current frame; calculating an average weighted neighborhood index value of the current frame in combination with the adjacent frame; setting a threshold value T; judging whether the t-frame tracking result is reliable or not according to a set threshold T; if the tracking is reliable, the tracker outputs a tracking result as a final tracking result; and if not, predicting a final tracking result by adopting an extended Kalman algorithm. According to the invention, the advanced tracker SiamRPN + + is taken as a base line, and lightweight improvement is carried out, so that the operation requirement of embedded equipment is met, and the speed and precision of the target tracker of the unmanned aerial vehicle are improved; a simple and effective index is also designed to detect shielding, and then target position prediction is realized by using extended Kalman filtering.

Description

Technical Field

[0001] The present invention belongs to, but is not limited to, the field of computer vision technology, and particularly relates to a lightweight anti-occlusion UAV target tracking method and system. Background Art

[0002] In complex environments, such as cities or dense forests, occluders may frequently appear, increasing the difficulty of UAV target tracking. Occlusion may cause the target to suddenly disappear from the UAV's field of view, resulting in the failure of the tracking method or misjudgment. In addition, the computing resources of the UAV platform are relatively limited, making it more difficult to implement real-time tracking methods.

[0003] To solve this problem, researchers have enhanced short-term trackers by introducing a tracking failure detection mechanism to make them suitable for occlusion scenarios. Generally, the detection of tracking failure can be divided into two categories. One is to design detection metrics using the attributes of the tracking response map, such as kurtosis, overlap ratio, confidence score, and peak sidelobe ratio. The other is to generate an offline network as a supervisor to evaluate the state of the tracker. Dai et al. used RTMDNet as the supervisor and set thresholds to determine whether to continue local tracking in the next frame. Wang et al. provided an online updated deep classifier for switching trackers and re-detectors. However, a single attribute is often not robust enough, and using a supervised network to evaluate the tracking state is time-consuming and computationally resource-intensive and not suitable for real-time tracking.

[0004] In the long-term tracking of UAVs, since the tracking of the current frame depends on the tracking state of the previous frame, when the target is completely occluded or disappears, the tracker inevitably loses the tracking target. And when the target reappears in the video, the tracker is unable to restart tracking. Most tracking methods solve this problem through a lost re-detection strategy. However, the re-detection process is complex and not suitable for real-time tracking. Based on the above analysis, the difficulties in applying target tracking to practical systems lie in achieving lightweight to ensure normal operation on embedded devices. On the other hand, it is how to design an efficient anti-occlusion scheme when encountering occlusion. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention provides a lightweight anti-occlusion UAV target tracking method and system for real-time tracking of UAV systems. The present invention is implemented as follows. A lightweight anti-occlusion UAV target tracking method includes the following steps: Step 1: Replace the backbone network of the target tracking method SiamRPN++ with the AlexNet network to obtain the lightweight target tracking method SiamRPN++_A.

[0006] Step 2: Obtain the video through the on-board camera, and select the target to be tracked in the video stream using bounding boxes.

[0007] Step 3: Run SiamRPN++_A to track the target in the video. For the current t-th frame, obtain the confidence map output by the classification branch. 。

[0008] Step 4: Calculate the peak value based on the confidence map output in the t-th frame. And the values within its k-neighborhood range. Calculate the weighted neighborhood index value of the t-th frame. 。

[0009] Step 5: Combine the first N historical frames of the t-th frame to calculate the average weighted neighborhood index value. 。

[0010] Step 6: Set the threshold T.

[0011] Step 7: If the average weighted neighborhood index value is greater than the set threshold T, it indicates that the tracking result is reliable. Then retain the tracking result output by the current tracker and continue tracking the next frame.

[0012] Step 8: If the average weighted neighborhood index value is less than the set threshold T, it indicates that the tracking fails. Then use the extended Kalman filter algorithm to determine the tracking result of the current t-th frame.

[0013] Furthermore, the specific steps to calculate the weighted neighborhood index value of the t-th frame in Step 4 are as follows: (1) where the function is to take the maximum value. function is to take the maximum value.

[0014] (2) Take the k values closest to the peak position, that is, obtain the values within the k-neighborhood range 。

[0015] (3) Calculate the weighted neighborhood index value of the t-th frame , , where is the balance weight coefficient, represents the th value ( , taking an integer) to the distance from the peak 。

[0016] Furthermore, the average weighted neighborhood index value in Step 5 is 。

[0017] Further, in step eight, the tracking result of the current t-th frame is predicted according to the prior estimation formula of the extended Kalman filter algorithm.

[0018] Another object of the present invention is to provide a lightweight anti-occlusion unmanned aerial vehicle (UAV) target tracking system for a lightweight anti-occlusion UAV target tracking method, including: A video image acquisition module that obtains video images through an on-board camera.

[0019] An input module that inputs the video into the tracking method SiamRPN++_A.

[0020] A tracking result acquisition module that runs SiamRPN++_A to track the target in the video. For the current t-th frame, a tracking result confidence map output by the classification branch is obtained. .

[0021] An average weighted neighborhood index value calculation module that calculates the weighted neighborhood index value for the current t-th frame and calculates the average weighted neighborhood index value of the current t-th frame in combination with adjacent frames.

[0022] A tracking result judgment module that sets a threshold T and determines whether the tracking result of the t-th frame is reliable according to the set threshold T.

[0023] An output module that, if the tracking is reliable, the tracker outputs the tracking result as the final tracking result; if it is unreliable, the extended Kalman algorithm is used to predict the final tracking result.

[0024] The tracking method SiamRPN++_A replaces the backbone network of the target tracking method SiamRPN++ with the AlexNet network.

[0025] Another object of the present invention is to provide a computer device. The computer device includes a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor executes the steps of the lightweight anti-occlusion UAV target tracking method.

[0026] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the lightweight anti-occlusion UAV target tracking method.

[0027] Another object of the present invention is to provide an information data processing terminal, and the information data processing terminal includes the lightweight anti-occlusion UAV target tracking system.

[0028] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present invention are: First, based on the advanced tracker SiamRPN++, this invention has made lightweight improvements to meet the operating requirements of embedded devices and improve the speed and accuracy of the UAV target tracker. In addition, for the occlusion problem, a simple and effective index is designed to detect occlusion, and then the extended Kalman filter is used to predict the target position.

[0029] Compared with the tracking technology using deep backbone networks, this invention adopts a shallow backbone network for lightweight tracking methods, effectively reducing the computational complexity of feature extraction and meeting the real-time requirements. Secondly, the tracking failure detection index designed in this invention combines the characteristics of the anchor box design strategy and uses the peak value of the confidence map of adjacent frames and its neighborhood information to robustly estimate the current state of the tracker in real time. Finally, this invention uses the extended Kalman filter algorithm to predict the flight trajectory of the UAV when occlusion occurs and has the ability to handle the state prediction of nonlinear systems.

[0030] Second, the expected benefits and commercial value after the transformation of the technical solution of this invention are as follows: 1. Improve the reliability and adaptability in UAV application scenarios Occlusion resistance: In complex environments, UAVs often face the situation where the target is occluded, and traditional tracking methods may fail. The occlusion resistance method you designed can effectively address this problem and improve the accuracy and stability of tracking. This can make UAVs more reliable when performing tasks in complex environments such as cities and forests.

[0031] Adapt to diverse application scenarios: From security monitoring, logistics distribution to search and rescue missions, UAVs need to continuously track targets in complex environments. Your technology can significantly improve performance in these scenarios, thus expanding the application scope and market demand of UAVs.

[0032] 2. Save resources and improve efficiency Lightweight algorithm: The lightweight design means that high tracking performance can be ensured even with limited hardware resources. This can reduce the dependence on high-performance computing platforms, thereby reducing costs, extending battery life, and improving the overall efficiency of UAVs.

[0033] Suitable for low-cost devices: Since high computing resources are not required, your system can be applied to relatively inexpensive hardware platforms, which enables the technology to be promoted to the consumer UAV market and increase the potential customer base.

[0034] 3. Expansion of application scenarios UAV automatic patrol and monitoring: The occlusion-resistant tracking technology can improve the performance of UAVs in security and monitoring. The demand for automatic patrol systems in enterprises, factories, farms, etc. is increasing day by day, and the ability to maintain continuous tracking in complex environments will be a huge market advantage.

[0035] Intelligent Logistics: In logistics distribution, especially in scenarios where drones need to autonomously avoid obstacles and track targets (such as delivery target addresses, vehicles, etc.), the anti-occlusion tracking system will greatly improve the accuracy and efficiency of distribution.

[0036] Search and Rescue Missions: When performing search and rescue missions in complex terrains such as mountains and forests, the anti-occlusion ability will significantly enhance the positioning and tracking capabilities of drones for search and rescue targets, increasing the success rate of search and rescue.

[0037] Third, the technical solution of the present invention fills the technical gaps at home and abroad in the industry: 1. Lightweight Anti-occlusion Target Tracking Algorithm Existing Problems: Existing anti-occlusion target tracking algorithms usually rely on deep learning models or complex computational methods, which require high-performance computing resources and limit their application on drone platforms with limited resources, especially on small, consumer-grade or high-endurance drones.

[0038] Filled Gap: The lightweight anti-occlusion method proposed by the present invention can obviously ensure good tracking performance even under limited computing resources. It fills the technical gap in achieving efficient anti-occlusion target tracking on low-power devices. This is particularly important for optimizing the target tracking system of small drones and is a direction that has received less attention at home and abroad.

[0039] 2. Target Tracking System Applicable to Complex Dynamic Environments Existing Problems: At home and abroad, most target tracking technologies still perform limitedly in dynamic scenarios (such as urban environments or complex natural environments), especially when the target is partially or completely occluded, they often cannot continue tracking. Although some academic research has proposed solutions, their engineering and practical applications are still not widespread.

[0040] Filled Gap: The solution proposed by the present invention is a lightweight solution that can effectively resist occlusion and continuously track targets in dynamic environments. This is particularly important in fields such as security, logistics, and rescue, filling the technical bottleneck in current actual application scenarios.

[0041] 3. Edge Computing and Low-latency Tracking Optimization Currently, the implementation of many anti-occlusion algorithms requires the assistance of high-performance computing platforms in the cloud for processing, which brings high network latency and power consumption problems. This may lead to unstable tracking in actual drone application scenarios, especially in tasks that require real-time decision-making, such as search and rescue, drone logistics, etc.

[0042] If the lightweight anti-occlusion technology proposed in the present invention can be efficiently implemented on edge devices (such as the drone itself) and can complete real-time tracking without relying on strong network support, it will fill the gap in the drone autonomous tracking technology in terms of low-latency and low-power edge computing.

[0043] Fourth, in the drone target tracking task, targets in complex environments are often occluded by trees, buildings, crowds, or other objects, making it difficult for the drone to continuously track. Existing target tracking algorithms may be effective under short-term occlusion, but when the occlusion lasts for a long time or covers a large area, target loss usually occurs. By designing a lightweight anti-occlusion algorithm, the technical solution proposed in the present invention can effectively handle occlusion problems in complex environments and maintain continuous tracking of the target. This breakthrough solves the problem that people have always hoped that drones can "see" occluded targets and quickly re-identify them after the occlusion is removed. Especially in the case of long-term occlusion or partial occlusion, this is a difficult point in the industry.

[0044] The computing power and battery life of drones are limited, which makes it difficult to apply many complex target tracking algorithms, especially large deep learning-based models that require a large amount of computing resources. The industry has been looking for efficient and accurate target tracking algorithms that can be implemented on power-constrained devices, but most rely on simplified and less accurate traditional algorithms. The lightweight anti-occlusion method proposed in the present invention can operate efficiently on hardware platforms with limited resources, breaking through the bottleneck of achieving high-precision target tracking on low-power devices in the industry. This not only extends the flight time of drones but also improves their tracking performance in actual tasks, especially suitable for small drones or low-cost drone platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flowchart of the lightweight anti-occlusion drone target tracking method provided by an embodiment of the present invention.

[0046] Figure 2 is a framework diagram of the lightweight anti-occlusion drone target tracking method provided by an embodiment of the present invention.

[0047] Figure 3 is a schematic diagram of the running speed and average distance precision rate results of different tracking methods provided by an embodiment of the present invention.

[0048] Figure 4 is a schematic diagram of the running speed and AUC score results of different tracking methods provided by an embodiment of the present invention.

[0049] Figure 5It is a schematic diagram showing the change of the average weighted neighborhood index value when target tracking enters the occlusion situation provided by an embodiment of the present invention; in part (a), one-third of the target car is occluded; in part (b), nearly one-half of the target car is occluded; in part (c), two-thirds of the target car is occluded; in part (d), the target car is completely occluded.

[0050] Figure 6 It is a structural diagram of a lightweight occlusion-resistant UAV target tracking system provided by an embodiment of the present invention.

[0051] Figure 7 The running trajectories of the car and the UAV and the predicted trajectory of the extended Kalman filter under the occlusion situation. Specific Embodiments

[0052] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] As Figure 1 shown, a lightweight occlusion-resistant UAV target tracking method provided by an embodiment of the present invention includes the following steps: (1) Obtain video images through an on-board camera.

[0054] (2) Input the video into the tracking method SiamRPN++_A.

[0055] (3) Obtain the confidence map of the tracking result.

[0056] (4) Calculate the weighted neighborhood index value for the current frame.

[0057] (5) Calculate the average weighted neighborhood index value of the current frame in combination with adjacent frames.

[0058] (6) Set a threshold T.

[0059] (7) Determine whether the tracking result of frame t is reliable according to the set threshold T.

[0060] (8) If the tracking is reliable, the tracker outputs the tracking result as the final tracking result.

[0061] (9) If it is not reliable, use the extended Kalman algorithm to predict the final tracking result.

[0062] The present invention proposes a lightweight occlusion-resistant UAV target tracking method, aiming to solve the occlusion problem in target tracking and improve the real-time tracking ability of UAVs. This method adopts a simplified neural network structure and an extended Kalman filtering algorithm to achieve robust tracking of targets under occlusion. The specific working principle is as follows: 1. Network replacement and lightweight model In the first step of the method, to achieve lightweight object tracking performance, the backbone network of SiamRPN++ is replaced with the AlexNet network to form a new lightweight model, SiamRPN++_A. Compared with the original SiamRPN++ network, AlexNet has a simpler structure and fewer parameters, so it can significantly reduce the computational load of the model, enabling the drone to perform real-time object tracking under limited computing power conditions, ensuring stability and efficiency during flight.

[0063] 2. Video acquisition and target selection The video stream is obtained in real time through the on-board camera carried by the drone, and the target is selected through the bounding box. The selected target will be continuously tracked in subsequent video frames. The bounding box determines the initial position of the target, which helps to identify and locate the target in each frame of the video stream, thus forming a continuous tracking effect.

[0064] 3. Confidence map generation and tracking result evaluation During the object tracking process, the SiamRPN++_A network classifies the target for each frame and generates a confidence map to represent the reliability of the object tracking result. The confidence map contains the probability distribution of the target's existence in the current frame. By extracting the peak value of the confidence map, the most likely position of the target can be obtained, and combined with multiple values in the neighborhood, the credibility of the tracking result can be further evaluated.

[0065] 4. Weighted neighborhood index calculation In the t-th frame, according to the peak value of the confidence map and the values in its surrounding neighborhood, the weighted neighborhood index value is calculated. This index calculates the tracking quality of this frame by weighting k points near the peak value and combining the balance weight coefficient and distance factor. This step can accurately reflect the reliability of the current tracking result through the peak neighborhood information when the target is occluded or the tracking effect is poor, providing a reference for subsequent frames.

[0066] 5. Average weighted neighborhood index and reliability judgment Combining the weighted neighborhood index of the t-th frame and the previous N frames, the average weighted neighborhood index value is calculated, and a threshold T is set for reliability judgment. If the average weighted neighborhood index value is higher than the set threshold T, it indicates that the tracking result is relatively stable, and the system will continue to use the current tracking result for the next frame of tracking; otherwise, the system considers that the tracking of the current frame fails, and thus initiates remedial measures to ensure the tracking effect even when the target is occluded.

[0067] 6. Prediction and compensation of the Extended Kalman Filter When it is determined that the tracking of the current frame fails, the system automatically calls the extended Kalman filter algorithm to predict the current position of the target through the prior estimation formula. The extended Kalman filter can smoothly predict the target position based on the position information of the previous few frames and correct the tracking deviation caused by occlusion. This method ensures that the UAV can still maintain stable target tracking in complex environments, making this method more practical and robust.

[0068] As Figure 2 shown, the selected target is made into a template frame, and the subsequent video frames to be detected are made into detection frames one by one. The template frame and the detection frame are input into the tracking method SiamRPN++_A. This algorithm adopts the shallow backbone network AlexNet and the lightweight tracking method depth network SiamRPN++, which effectively reduces the computational complexity of feature extraction and meets the real-time requirement. Secondly, it is proposed to detect the target occlusion situation that occurs during the process of the UAV tracking a moving target according to the confidence map output by the classification branch. If the tracking fails, the extended Kalman filter algorithm is used to predict the tracking result, otherwise the result output by the tracking method is used as the tracking result.

[0069] Embodiment 1 A lightweight occlusion-resistant UAV target tracking method includes: (1) Using a quadrotor UAV to capture a video.

[0070] (2) Transmitting the image to the embedded device Nvidia Jetson through the video link.

[0071] (3) Nvidia Jetson runs the lightweight target tracking method SiamRPN++_A and manually selects the bounding box of the target.

[0072] (4) Running SiamRPN++_A to track the target in the video. For the current t-th frame, obtain the tracking result confidence map output by the classification branch .

[0073] (5) Calculate the peak value and the values within its k-neighborhood range , where k is taken as 8, and calculate the weighted neighborhood index value of the t-th frame. The formula is as follows: , where The function is to take the maximum value.

[0074] Take the peak The k values closest to the position, that is, obtain the values within the k-neighborhood range .

[0075] Calculate the weighted neighborhood index value of the t-th frame , , where is the balance weight coefficient, represents the th value ( , taking an integer) to the peak . Here .

[0076] (6) Combine the first N historical frames of the t-th frame, where N is taken as 6, and calculate the average weighted neighborhood index value , and the formula is as follows: .

[0077] (7) Set the threshold T = 7.

[0078] (8) If the average weighted neighborhood index value is greater than the set threshold T, it indicates that the tracking result is reliable, then retain the tracking result output by the current tracker and continue to track the next frame.

[0079] (9) If the average weighted neighborhood index value is less than the set threshold T, it indicates that the tracking fails, then use the extended Kalman filter algorithm to determine the tracking result of the current t-th frame.

[0080] Embodiment 2 Use the dataset DTB70 for experiments. Use the embedded device NVIDIA Jetson to compare the speed and accuracy of SiamRPN++_A and other tracking methods. Include SiamRPN++_M, DaSiamRPN, Ocean, LightTrack, Se-SiamFC, SiamAPN, SiamBAN, SiamCAR, SiamFC++, SiamGAT, SiamMask, UpdateNet, SiamSA-IROS and SiamSA-TII. Figure 3 Shows the results of the running speed and average distance precision of different tracking methods. Figure 4 shows the results of the running speed and AUC score (success) of different tracking methods. SiamAPN and SiamRPN++_A can run at a speed of more than 30 FPS on NVIDIA Jetson and meet the real-time requirements. SiamRPN++_A performs well in terms of average distance precision and AUC score metrics, and meets the UAV speed requirements at the same time.

[0081] As Figure 5As shown in part (a), when one-third of the target vehicle is occluded, the extracted target features are still discriminative. At this time, the average weighted neighborhood index value calculated based on the tracking result confidence map is 7.9976, which is greater than the threshold. Therefore, the algorithm determines that it is not occluded. Similarly, when nearly half of the vehicle is occluded, as Figure 5 shown in part (b), the average weighted neighborhood index value at this time decreases, but is still greater than the threshold. Therefore, it still belongs to unoccluded. When two-thirds of the vehicle is occluded, as Figure 5 shown in part (c), the average weighted neighborhood index value is 0.6696, which is less than the threshold and is determined to be occluded. When the target vehicle is completely occluded, as Figure 5 shown in part (d), the peak value of the tracking result confidence map is extremely low and there are multiple peaks, and the average weighted neighborhood index value is 0.0272. Through the above analysis, it can be seen that the proposed tracking failure detection index can effectively reflect the target occlusion situation and provide an effective reference for the tracking method to detect occlusion. And during occlusion, the extended Kalman filter can basically determine the target trajectory.

[0082] Specific application fields or related products of the present invention.

[0083] 1. Security monitoring field Application scenario: In scenarios such as urban security, enterprise security, and border monitoring, drones can monitor targets through aerial patrols. Traditional surveillance cameras are easily restricted by the field of view, while drones have flexible movement capabilities and can cover a larger area. The anti-occlusion technology ensures that even in the complex urban environment with buildings and trees blocking, it can still continuously track suspicious personnel or vehicles.

[0084] Related products: Drone urban patrol monitoring system.

[0085] Industrial park or factory area drone security system.

[0086] Border and coastline patrol drones.

[0087] 2. Intelligent logistics distribution Application scenario: Drone logistics distribution needs to perform autonomous navigation and target tracking in a complex urban environment. The anti-occlusion target tracking technology can help drones continue to identify the target location in case of occlusion, such as tracking delivery vehicles and package delivery locations, improving the accuracy and efficiency of drone delivery.

[0088] Related products: Automatic drone delivery system (such as food delivery, goods delivery).

[0089] Intelligent UAV Scheduling System for Warehouses and Distribution Centers

[0090] 3. Search and Rescue and Emergency Response Application Scenario: In complex environments such as disaster sites, forests, and mountains, anti-occlusion technology can help UAVs maintain effective monitoring of missing persons or target areas during search and rescue missions. Even in the presence of occlusions such as trees and rocks, it can improve the search efficiency and success rate. UAVs can quickly cover large areas, reducing the time and risk of manual search and rescue.

[0091] Related Products: Mountain and Forest Search and Rescue UAV System

[0092] UAV Search and Rescue and Reconnaissance System after Natural Disasters (such as earthquakes and floods)

[0093] Maritime Search and Rescue UAV

[0094] 4. Cooperative Operations between Driverless Cars and UAVs Application Scenario: UAVs can work in cooperation with driverless cars to provide aerial tracking and navigation assistance. Especially in complex urban traffic environments, anti-occlusion technology can help UAVs maintain tracking even when the car is occluded (such as passing through tunnels or buildings), enabling all-round monitoring and cooperative operations.

[0095] Related Products: Cooperative Navigation System for Driverless Cars and UAVs

[0096] UAVs in Traffic Monitoring and Management Systems

[0097] 5. Agriculture and Environmental Monitoring Application Scenario: In the agricultural field, UAVs can be used for crop monitoring, pest and disease detection, and irrigation management. Anti-occlusion technology can help UAVs effectively track the growth of crops in dense vegetation or farmland. Similarly, in environmental monitoring, UAVs can traverse complex terrains (such as forests or wetlands) to continuously track specific animals and plants or monitor pollution sources.

[0098] Related Products: Intelligent Farmland Monitoring UAV

[0099] UAV Monitoring System for Environmental Pollution

[0100] Forest Health Monitoring UAV

[0101] When occlusion occurs, it is necessary to take measures to avoid tracking failure. In this experiment, the Extended Kalman Filter algorithm is used to predict the position of the target at the next moment, and then this position information is sent to the flight control system of the UAV for PID control. According to the UAV trajectory data calculated by the optical infrared motion capture system, when the tracking is correct in the early stage, the tracking result is used as the observation value to continuously update the parameters of the Extended Kalman Filter algorithm. When occlusion is detected, at this time, the UAV predicts the position of the UAV at the next moment according to the historical trajectory and the Extended Kalman Filter algorithm. As shown in Figure 6, the black line is the running trajectory of the car without occlusion, and the trajectory of the car in the occluded part is represented by a blue dotted line. The red dotted line is the trajectory of the UAV flying according to the output result of the tracking algorithm, indicating that the tracking result is reliable during this period. When the car encounters occlusion, at this time, the tracking result is unreliable, and the Extended Kalman Filter (EKF) is used for trajectory prediction, and the prediction result is shown by the green dotted line. It can be observed that the Extended Kalman Filter can roughly predict the route that the UAV should travel when encountering occlusion. Figure 7 The running trajectories of the car and the UAV and the predicted trajectory of the Extended Kalman Filter under occlusion.

[0102] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, such as provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable logic devices such as field programmable gate arrays, or can be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software, such as firmware.

[0103] The above is only the specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A lightweight anti-occlusion UAV target tracking method, characterized in that: The following steps are involved: Step 1: Use the AlexNet network to replace the backbone network of the target tracking method SiamRPN++ to obtain the lightweight target tracking method SiamRPN++_A; Step 2: Obtain video through the onboard camera and use the bounding box to select the target to be tracked in the video stream; Step 3: Run SiamRPN++_A to track the target in the video. For the current tth frame, obtain the tracking result confidence map output by the classification branch. ; Step 4: Calculate the peak value based on the confidence map output from the tth frame and its k-neighborhood value , calculate the weighted neighborhood index value of the tth frame ; Step 5: Combine the previous N historical frames of the tth frame and calculate the average weighted neighborhood index value ; Step 6: Set the threshold T; Step 7: If the average weighted neighborhood index value If it is greater than the set threshold T, it indicates that the tracking result is reliable, then the tracking result output by the current tracker is retained and the next frame tracking is continued; Step 8: If the average weighted neighborhood index value If it is less than the set threshold T, it indicates that the tracking fails, and the extended Kalman filter algorithm is used to determine the tracking result of the current t frame.

2. The lightweight anti-occlusion UAV target tracking method according to claim 1, characterized in that: In step 4, the weighted neighborhood index value of the tth frame is calculated The specific steps are as follows: (1) ,in The function is to take the maximum value; (2) Take the peak value The k nearest values, that is, the values ​​within the k neighborhood ; (3) Calculate the weighted neighborhood index value of the tth frame , ,in is the balance weight coefficient, Representative Value to peak distance.

3. The lightweight anti-occlusion UAV target tracking method according to claim 1, characterized in that: Average weighted neighborhood index value in step 5 .

4. The lightweight anti-occlusion UAV target tracking method according to claim 1, characterized in that: In step eight, the current t-frame tracking result is predicted according to the prior estimation formula of the extended Kalman filter algorithm.

5. A system for a lightweight anti-occlusion UAV target tracking method as claimed in any one of claims 1 to 4, characterized in that: include: A video image acquisition module, which acquires video images through an onboard camera; Input module, which inputs the video into the tracking method SiamRPN++_A; The confidence map acquisition module runs SiamRPN++_A to track the target in the video. For the current t-th frame, the confidence map of the tracking result output by the classification branch is obtained. ; An average weighted neighborhood index value calculation module calculates a weighted neighborhood index value for a current frame and calculates an average weighted neighborhood index value for the current frame in combination with adjacent frames; The tracking result reliability judgment module sets a threshold T and determines whether the tracking result of frame t is reliable according to the set threshold T; Output module, if the tracking is reliable, the tracker outputs the tracking result as the final tracking result; If it is unreliable, the extended Kalman algorithm is used to predict the final tracking result; The tracking method SiamRPN++_A adopts the AlexNet network to replace the backbone network of the target tracking method SiamRPN++.

6. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the lightweight anti-occlusion UAV target tracking method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the lightweight anti-occlusion UAV target tracking method as described in any one of claims 1-4.

8. An information data processing terminal, characterized in that: The information data processing terminal includes the lightweight anti-occlusion UAV target tracking system as described in claim 5.

Images