A Visual Positioning Algorithm For Drones Based On Deep Learning

With the gradual maturity of drone technology,more and more fields have introduced drones to undertake work.UAVs have low operating costs and strong environmental adaptability,they are suitable for use in dangerous conditions such as disaster relief and forest fire fighting.The premise that drones can perform air missions is to accurately locate its own position.Traditional navigation technologies mainly rely on Global Positioning System(GPS)or inertial navigation systems.However,the global positioning system is limited by the working environment and it is controlled by the United States,the reliability and security in the defense field are poor meanwhile the inertial navigation system cannot maintain accurate positioning for a long time due to its accumulated errors.The use of visual perception to assist the positioning and navigation of drones has become a common navigation method.However,for the need to specifically design landmarks,the existing visual positioning methods are less robust.Therefore,it is important to study a new UAV visual positioning method.This thesis studies a UAV visual positioning method based on deep learning and PnP technology.The research contents are as follows:(1)For the possible distortion of the camera,it is calibrated with the camera model to obtain relevant internal parameters and distortion coefficients to eliminate errors caused by distortion.(2)Aiming at the high requirements of the current visual positioning algorithms for landmarks,the two models of YOLOv2 and Faster RCNN are trained to perform landmark detection tasks,which realizes more robust identification and localization of 2D landmark key points in the image.For the poor detection effect of YOLOv2,the improvement is made by modifying the loss function.In view of the inconsistent detection effect for targets of different scales in Faster RCNN,it is improved by combining multi-scale training strategies.By comparing the detection performance of improved YOLOv2 and Faster RCNN,the advantages of improved Faster RCNN model are illustrated.(3)In order to accelerate the speed of drone visual positioning,a simplified Faster RCNN model is designed with the help of transfer learning.Compared with the unimproved,the simplified model has no significant difference in the detection effect of the landmarks but the detection speed has been greatly improved.(4)Finally,according to the detection result of landmark images,a positioning algorithm based on object detection and PnP technology is designed.The algorithm uses the detected landmark coordinates in image and corresponding world coordinates to solve the UAV position.Through the indoor flight experiment of the UAV,the feasibility of the algorithm is verified,and the positioning accuracy meets the expected requirements.