Research On Key Technology Of Object Detection,tracking And Visual Servo Based On UAV

In order to enable UAVs to detect hidden targets in forest,indoor and complex city with weak or even invalid GPS signals and take effective measures in time,it is necessary to equip UAVs with visual sensors and other mission equipment to collect more reliable image information of moving targets and detect,identify,and track them.At the same time,considering the short closed-loop cycle of UAV control system,it is necessary to ensure the real-time requirement of each link of the vision system and the accuracy and stability of the whole visual servo system.Based on the visual image processing algorithm and the Four-rotor flight control principle,this paper studied the target detection,target recognition,target tracking and visual servo technology,and designed the small Four-rotor UAV vision recognition and tracking system.The specific research work and results are as follows:Firstly,this paper implemented the MobileNet-SSD algorithm on the embedded task board.In the aspect of target detection and recognition,based on the SSD target detection algorithm,this paper processed the video stream collected by the airborne camera to realize real-time detection of ground moving targets in the UAV embedded task board.At the same time,by combining the more lightweight MobileNet and SSD,the real-time performance of the SSD algorithm is improved.Secondly,for embedded devices with poor computing performance,this paper improved the TLD algorithm.In the aspect of target tracking,based on the classic long-term target tracking algorithm TLD,the improved strategy is adopted by narrowing the target search area and dynamically adjusting the variance classifier threshold.Compared to the native TLD algorithm,its real-time performance is improved by about 10%,while its accuracy has been slightly improved.Then,for embedded devices with strong computing performance,this paper improved the CFnet algorithm.By studying SiamFC based on deep learning and CFNet which combines the characteristics of related filters,this paper designed a dual full convolution twinning network to improve the robustness of visual target tracking by combining the appearance features in the similarity learning problem with the semantic features in the image classification problem.Finally,this paper designed a visual servo control system for drone.This paper discussed the selection of the hardware platform of the visual recognition and tracking system of the UAV and the implementation of the software system.Using the PID control law,the visual servo IBVS control program is designed to realize the real-time tracking of the drone to the selection target.