Research And Implementation Of Obstacle Detection Based On Binocular Vision

Real-time environment perception is one of the important capabilities that autonomous mobile robots need to possess.Relying on various sensors,the robot can obtain information from its environment.Sensors used to obtain 3D scene information include lidar,structured light camera,TOF camera and binocular camera.Compared with other sensors,binocular vision sensors are cheaper and can meet most needs when matched with appropriate algorithms.Based on this premise,this thesis will design and implement a set of obstacle detection systems using binocular vision sensors,which require accurate real-time detection of obstacles in the scene.The main work and innovations of this thesis are as follows:First,establish the corresponding mathematical model for the monocular camera imaging principle and the binocular camera measurement principle to provide theoretical support for the subsequent algorithms.Each parameter in the model needs the corresponding calibration algorithm calculation,so the calibration algorithm and stereo calibration algorithm principle of commonly used pinhole cameras are studied.After performing stereo correction on the collected image,the result obtained can be regarded as an ideal binocular camera collection.Secondly,the realization principles of three types of commonly used stereo matching algorithms are studied separately,because the stereo matching algorithm is mainly used to calculate disparity data from binocular images,which is a key step in restoring three-dimensional information.In the experiment process,starting from the theory,combined with the existing results to eliminate the global algorithm that cannot be matched in real time.After comparing the matching effect of the semi-global algorithm and the local algorithm in the actual scene,the former is finally selected,and a simple experiment is designed to verify the accuracy of the selected algorithm.Finally,it is found that the matching effect of the semi-global stereo matching algorithm in a certain range basically meets the demand.Thirdly,this article completes obstacle detection based on disparity data,because the disparity map can be used to restore the distribution information of threedimensional points in the scene.Obstacle detection generally uses the camera external parameters,using the characteristics of the ground parallax when a specific camera is installed,this thesis uses the RANSAC algorithm to dynamically estimate the camera external parameters from the v-disparity map in real time.In order to detect obstacles more robustly,after comparing the methods proposed in multiple documents,the MultiLayer Stixel algorithm is finally selected to complete the segmentation of free space and obstacles.This method can effectively reduce the problem of noise interference in disparity data.A good estimate of the road surface,obstacles on the road surface and even suspended objects in the scene.Aiming at the problem that the detection results of the Multi-Layer Stixel algorithm are not related,this thesis proposes a method of clustering the detection results using DBSCAN.After clustering,the obstacle segmentation can be completed on the image.Finally,this thesis designs and implements a simple mobile platform for testing the overall effect of the obstacle detection system.The mobile platform is equipped with cheap binocular cameras and uses the obstacle detection system designed and implemented in this thesis.In order to allow the car to move autonomously,a simple navigation algorithm is also designed for it.Then in the experimental environment,it is tested whether the platform can accurately calculate the position of obstacles in real time,while continuing to move autonomously in the experimental scene.Experimental results show that the system can detect most obstacles,and the speed of the detection system to process binocular images meets real-time requirements.