Research On Obstacle Avoidance Technology Of Rotor UAV Based On Monocular Vision

UAVs are widely used in civil and military fields because of their advantages such as simple structure,flexible flight,easy control,and low cost.In recent years,intelligence is a development trend of UAVs while obstacle avoidance is the problem that must be solved.This thesis aiming at the subject of autonomous obstacle avoidance for multi-rotor UAVs in outdoor environments,controlling the UAV to avoid obstacles that identified from the images obtained by a camera installed in front of the UAV according to an obstacle avoidance strategy.The main work of this thesis is as follows:The ORB algorithm has the advantages of small calculation amount,fast calculation speed,and rotation invariance.Therefor,it has been chosen and improved to possess scale invariance,which is in order to meet the needs of this thesis,reduce calculation amount and increase obstacle identifying speed.At first,the original image is reduced based on the bilinear interpolation method,and multiple reduced images with different reduction coefficients are constructed into a scale space pyramid,the scale of each group is the reduced coefficient,which become a linear growth relationship.Then extract the FAST feature points in each group of the pyramid,sort the feature points according to FAST response value and keep the fist 2n,after that sort the 2n points according to Harris response value and keep the first n as the feature points of the group.At last calculate the BRIEF descriptor in each group.This thesis uses brute force matching based on Hamming similarity to match the feature points of two images and eliminates the mismatch based on the Euclidean distance and the scale of the group where the feature points belong.The performance of the improved ORB algorithm is verified by experiment and comparison of the effects of SIFT algorithm and SURF algorithm.In terms of obstacle detection,this thesis establishes a relationship between scale ratio of objects in two images and distance,then calculates the scale ratio of each matching feature point,selects the feature points corresponding to the obstacle according to the relationship,at last obtains the range of obstacle according to the distribution of them.In terms of obstacle avoidance strategies,this thesis judges whether the obstacle needs to be avoided and direction of avoidance based on the range of the obstacle and the row in image in front of the UAV,at last establishes a relationship between the physical width and the pixel width to calculate the distance to avoid obstacles.Finally,this thesis verifies the feasibility of the algorithm through 3D modeling in ROS system.The experimental results show that the algorithm in this thesis has a good recognition effect when the obstacle is not blocked.The time from acquiring the image to determining whether to avoid obstacles is only 50 ? 120 ms,which can fully meet the real-time requirements of monocular visual obstacle avoidance.