Research And Implementation Of Autonomous Landing System For UAV Based On Computer Vision

UAVs usually use a combination of INS and GPS to achieve navigation.However,the accuracy of INS and GPS is relatively limited,so that it is difficult to achieve accurate autonomous landing with them.In recent years,with the development of computer vision and micro-processors,using vision to assist UAVs in autonomous landing has become a feasible solution.In this paper,an autonomous landing system for UAVs is designed by comprehensively using technologies such as target detection and flight control.At present,object detection technologies generally require a large number of image calculations,while the computing power of the image processing chip of the general UAV is relatively limited.In this paper,an efficient landmark detection module is designed to reduce the amount of calculation for detection.The detection system includes three independent modes: low-power mode,which uses a contour-based detection scheme;high-performance mode,which uses a YOLOv4-Tiny-based detection scheme;hybrid mode,which combines the previous two modes and switches between the two detection schemes.In the contour-based detection scheme,image preprocessing methods such as filtering and edge processing are performed on the collected images to obtain the contour information of the landmark,so as to complete the detection according to the contour feature of the landmark.In the detection scheme based on YOLOv4-Tiny,YOLOv4-Tiny is used to train,and the feature channels and parameters are trimmed to improve the efficiency of algorithm execution,and then the trained model is used to complete the detection.In contrast,the former has a smaller amount of calculation and faster detection speed,but lower detection accuracy,which is suitable for conventional landing scenes;while the latter has a larger calculation amount and slower detection speed,but has higher detection accuracy and is suitable for complex landing scenes.Experiments have verified that both detection schemes can achieve good results on UAV platforms with limited computing power,but they have their own advantages and disadvantages in detection speed and detection accuracy.After the landmarks are recognized,Kalman filter is used to fuse image data and IMU data to correct the attitude of the UAV,thereby assisting the UAV to achieve precise autonomous landing.