Research On Visual Navigation Based On Quadrotor

With the deepening of machine vision research,vision-based tracking and navigation has been widely applied to various important fields,such as driverless,intelligent robots,military strikes and so on.This thesis takes the seventh generation mission of the International Air Robot Competition(IARC)as the background of the subject,and uses the quadrotor as a platform to study the algorithms such as monocular visual odometer,boundary detection and target location tracking for realizing the mission.First,in order to develop a corresponding flight strategy for the aircraft,the system needs to know the position of the target and the aircraft in the field,ie their world coordinates.However,the image acquired by the monocular camera is a two-dimensional image coordinate.For this purpose,a monocular visual odometer based on the principle of camera calibration and inverse perspective mapping is proposed,which realizes the global positioning of the target.At the same time,in order to make the strategy more effective,the ground robot is modeled and analyzed,and the target motion area is roughly predicted.Secondly,this paper combines computer vision and UAV inertial navigation to study and analyze the tracking algorithm.Considering the requirements of robustness and real-time in the game and the limitation of computing capacity of the onboard computer,The improved Camshift algorithm(continuous adaptive Meanshift algorithm)detects and tracks the target.For the algorithm,the target needs to be manually selected,and the target template is introduced to realize the automatic tracking of the algorithm.For the target occlusion situation that may occur in the actual game,the Kalman filter is combined in the tracking algorithm to predict the target after occlusion.The single target tracking method achieves multi-target tracking by combining background segmentation.Finally,in order to meet the requirements of the game that the aircraft must not fly out of the boundary for more than 5 seconds,it is necessary to perform high-precision detection on the boundary of the game and return in time.Because the boundary features of the competition venue are complex,the traditional feature detection method is difficult to meet the requirements and the failure rate is high.In order to solve this problem,a boundary detection based on the support vector machine(SVM)with direction gradient histogram(HOG)is designed.And the performance of the detector is improved by bootstrap method and cross-validation method,and the detector has a good Lu scale for the scale,rotation and illumination of the boundary of the site.Great.