Exposure Control And Obstacle Detection For UAV Based On Stereo Vision

Recent improvements of micro multi-rotors have enabled various commercial applications,including aerial photograph,power line inception,and precision agriculture.At the same time,the complex environment faced by different tasks puts higher demands on the environmental awareness of drones.Stereo vision system simulates the way the human eye observes the world through two cameras in parallel.Compared with active sensors such as Lidar,TOF,and ultrasonic,visual sensing is crucial due its relatively low cost and high information throughput.There are still many research hotspots for drones obstacle avoidance and navigation,considering the limited computing power of the onboard processor and the complexity of the environment.For instance,how to effectively control camera exposure in a highly dynamic environment to capture detailed image data,and how to estimate disparity through fast stereo matching and effectively perform obstacle detection.This dissertation mainly designs a UAV stereo vision perception system for the perception of surrounding obstacles.The main work is as follows:1)Camera automatic exposure algorithm based on effective brightness of image feature points position.The image acquired by the visual sensor on the drone will be directly fed into followed algorithms as input.So the object features,details and other information in the image will have a great impact on the accuracy of the algorithms.However,image acquisition with high exposure quality is often limited by sensor and algorithms.The camera automatic exposure algorithm based on the effective brightness of image feature points designed in this dissertation can fully consider the exposure quality of the feature point rich area.So it can quickly iterate to the proper exposure in high dynamic environment to capture important image features.2)Fast stereo matching and obstacle perception.Stereo vision system obtains depth infonnation by simulating the way humans observe the world and using the disparity principle to match the left and right images.Due to the limited load and power of the drone,the algorithms need to take care of both accuracy and speed.The stereo vision system designed in this dissertation implements the CUDA-accelerated semi-global matching algorithm on the NVIDIA Jetson TX2 embedded processor,and then performs depth reconstruction based on the disparity map.Finally,efficient perception and representation of obstacles are performed in the generated octree map.In this dissertation,the accelerated end-to-end neural network model is used for disparity estimation in complex cases.