Research Of Obstacle Avoidance Technology Based On Machine Vision For Unmanned Aerial Vehicles

Multi-rotor UAV construction technology is relatively simple and easy to operate and control.UAV plays an important role in aerial photography,rescue search,agricultural plant protection,earthquake relief,geographic mapping and other fields.The autonomous obstacle avoidance function can ensure that drones can complete complex and difficult tasks.With the development of machine vision technology,monocular vision technology has gradually been applied to the obstacle avoidance of drones.Domestic and foreign research in this area has made great breakthroughs.This thesis focuses on the problems of monocular vision technology in the processing strength,obstacle recognition accuracy and obstacle avoidance path decision,and designs a monocular obstacle avoidance system that separates image processing from flight control systems,and uses a reactive avoidance obstacle strategy with the help of Lidar.In order to meet the needs of the obstacle avoidance system for the monocular camera parameters,the principle of Zhang Zhengyou's calibration is studied,and it is verified through experiments that the calibration accuracy of the calibration algorithm can meet the system requirements.The movement of the drone can cause the camera perspective to change.Changes in image rotation,size,and lighting affect the stability of the feature points.In order to improve the robustness of feature points,this thesis studies two kinds of important algorithms based on multi-scale local invariant feature algorithms: SIFT algorithm and SURF algorithm.Through experiments,the performance differences between SIFT algorithm and SURF algorithm are compared and analyzed.This thesis uses the image matching algorithm based on local features-FLANN algorithm to match the features of the two images.According to the Euclidean distance of feature points and the change of the feature point scale,the matching items are optimized.In addition,according to the relative size clues,the relationship between the scale ratio of the feature points and the depth of the image in the matching pair is studied.Through the experiment,the feasibility of extracting the obstacle information from the algorithm is verified.In pure monocular visual obstacle avoidance,the obstacle avoidance algorithm may fail due to insufficient image information.This thesis introduces the low-cost lidar data as a supplement to visual obstacle avoidance.According to the relationship between image pixel points and Lidar scanning points,the obstacle region information is integrated with the distance data to make the depth information of the obstacle precise.In the obstacle avoidance path decision-making,in order to be able to avoid the most dangerous obstacles in real time,a safety window is designed based on the drone size and the safety distance of drone flight.The nearest obstacle in the security window is the preferred evasion target,and the fastest direction of dodge is evaluated based on the regional search to complete the avoidance.The thesis has simulated the research content to verify the feasibility of the monocular obstacle avoidance system function and algorithm.The image processing function runs as a server software on a desktop computer.This thesis builds a drone three-dimensional simulation platform based on the ROS open source operating system.Obstacle avoidance path decision function is implemented and operates as a client on a drone simulation platform.The wireless communication between the client and the server is realized so that the entire monocular obstacle avoidance system can operate normally.Through the data analysis of the experimental results,the obstacle avoidance system has better performance when the drone flight speed is between 1m/s and 8m/s,and the surrounding environment is relatively simple.