Research On Road Target Detection And Ranging System Based On Binocular Vision

With the development of economy and the progress of science and technology,the number of cars is increasing,resulting in a more complex traffic environment.The driving safety problem is increasingly serious.Perception of the car’s surroundings is key to solving this problem.The environment perception system based on binocular vision has attracted the research interest of scholars from all walks of life because of its relatively simple structure,low construction cost,and can assist or even replace the driver’s driving.After in-depth study of binocular ranging and target detection technology,this thesis builds a target detection and ranging system based on binocular vision and depth learning to verifies the feasibility of this scheme.The details are as follows:(1)Binocular ranging part: The imaging principle of the monocular camera and the ranging principle of the binocular camera are analyzed by the geometric relation.The binocular camera used in this thesis is calibrated using the Zhang Zhengyou calibration method to obtain the internal and external parameters of the two monocular cameras and the spatial relationship matrix between the binocular cameras;Distortion correction and stereo correction are performed on the binocular images.In terms of stereo matching,compare and analyze the disparity maps generated by SGBM(Semi Global Block Matching)and BM(Block Matching)algorithms.In combination with the requirements of the use scenarios in this thesis,BM algorithm is used as the stereo matching algorithm to calculate the disparity in the shortest calculation time,and the actual distance is obtained according to the disparity rotation distance formula.(2)Target detection part: Build the data set required for the three types of road common targets to be detected in this thesis.Due to the overall size of the car category and the presence of many weak texture regions,it is easy to generate mismatches when calculating disparity.In order to accurately perceive the position of the vehicle,this thesis approximates the distance of the license plate area to the distance of the vehicle.Therefore,this thesis also integrates the domestic license plate data set,and uses Label Img software to re-label the image.Finally,YOLOv4 target detection model is built then train and test it.The rationality of the model used in this thesis is proved by comparing the average accuracy rate,average accuracy rate and detection frame rate of the detection categories with the mainstream target detection models.(3)System test part: Under the computer hardware platform,use Py Qt software to integrate the binocular ranging and target detection functions into the graphical interactive interface of the system,and conduct the actual test of the whole system at multiple sites and distances.The experimental results show that the target ranging error is less than 5% in the range of 0.5m to 7.5m.