Research On Vehicle Ranging Obstacle Avoidance And Target Recognition Based On Binocular Stereo Vision

With the development of China’s economy and society and the advancement of science and technology industry,cars have gradually entered more families’ lives and work.But with the road congestion and frequent traffic accidents,active safety of automobiles has become an important issue facing the development of automotive technology.In this paper,the environment perception system,one of the key modules of the autopilot system,is deeply studied.Based on the analysis of the advantages and disadvantages and the research status of the sensor technology commonly used in automotive environment perception,an automobile ranging obstacle avoidance and target recognition system based on binocular stereo vision is proposed.Firstly,the mathematical principle of camera imaging and binocular stereo vision ranging is studied and analyzed.Combined with the requirements of binocular ranging system,a suitable visual sensor is selected to build a binocular stereo vision platform to obtain binocular images.Zhang Zhengyou’s plane calibration method is adopted.The binocular camera is used to capture the left and right views of the checkerboard at different angles.The internal and external parameters of the binocular camera are calibrated based on the MATLAB calibration toolbox.According to the matching requirement of the binocular image by the stereo matching algorithm,the binocular image is grayed out and filtered and denoised in advance,and the binocular camera parameters obtained by calibration are used to eliminate the distortion and stereo correction of the left and right views to obtain a strictly aligned binocular image.Then,through the analysis of the stereo matching algorithm,the gray of the image region is extracted as the matching primitive.Under the various constraint criteria,the local optimal search strategy is used to determine the matching point of the binocular image with the absolute error accumulation function to obtain the matching point.The left and right parallax calculates the depth information of the target.Secondly,according to the requirements of real-time identification and ranging of vehicles,select the appropriate real-time deep learning algorithm,train based on the road scene picture data set,identify the common cars,bicycles,pedestrians,etc.in the road scene,and output the target area coordinates.Obtain the target distance in the stereo matching disparity map.Finally,the accuracy,real-time and feasibility of the binocular stereo vision system were verified by three experimental schemes: static,dynamic and portable.The static ranging experiment performs binocular ranging and analysis error on fixed targets at different distances.The dynamic target ranging experiment detects the moving target in front and acquires the distance by the target recognition algorithm in real time.Portable real vehicle test ported the binocular stereo vision system to the NVIDIA embedded board,and controlled the test vehicle in the road scene to achieve effective ranging,identification and automatic obstacle avoidance functions for the target.