Car Type Recognition And Body Modeling Based On Binocular Stereo Vision

The current car body modeling methods mainly include modeling software method,laser scanner method and three-coordinate measuring method.Modeling software has great limitations on modeling complex surfaces.Laser scanner is expensive and needs a large amount of date.The three-coordinate measuring instrument is easy to scratch the body's paint surface when modeling,and is easy to be restricted by space because of its bulk.With the development of computer vision technology,3D modeling technology based on binocular stereo vision has become a new research hotspots due to its lower cost and higher efficiency.Therefore,this paper uses binocular stereo vision technology to model the car body,and builds the car body model database with the car type recognition technology to provide model data support for car wash,vehicle beauty,equipment processing and other applications.The paper's main research contents:?1?The research of the camera calibration technology.After comparing the characteristic of the current calibration algorithm,the camera is calibrated by using the checkerboard calibration method.The calibration algorithm is used to obtain the internal and external parameters of the camera and the relative position of the binocular camera,so as to eliminate the image distortion and rectify the epipolar of images.?2?The research of the car detection and car type recognition technology.The SSD object detection algorithm is adopt in car detection.According to the complex characters of the car,the VGG₁6 network has improved,and the multi-scale and multi-feedback deep-revolution neural network car model recognition model has designed.This model solves the shortcoming that VGG₁6 can only input a single size image,resulting in poor feature extraction.The final experiment proves that the algorithm has good performance in car type recognition.?3?The research of the stereo matching algorithm.An improved Minimum Spanning Tree?MST?stereo matching algorithm is proposed.This algorithm combines the disparity range estimation to calculate fast the matching cost reducing the computational of the matching cost,and introduces the gradient information to improve the sensitivity to pixelchanges in low-texture regions.Finally,Middlebury data and real scene data are used to verify the effectiveness of the algorithm.?4?Design of car type recognition and modeling experiments.The key steps of the experiments are image acquisition,image correction,car type recognition,stereo matching,and body modeling and so on.And they are demonstrated and verified.By experiments,the feasibility and validity of car type recognition and body modeling based on binocular stereo vision are verified.