Research Of Real-time 3D Reconstruction And Measurement System Based On Binocular Stereovision

3D reconstruction,which combined with image processing,stereo vision,multiple view geometry and computer graphics,is an important research direction in computer vision field.Image based 3D reconstruction technology has rich applications in fields of video surveillance,VR,mapping and navigation and etc,because of its advantages of low cost and abundant texture.With the development of image sensor technology,the resolution of camera becomes higher and higher.More image information makes the result of 3D reconstruction more accurate,but it also brings challenges to real-time reconstruction.This paper investigates core links for constructing a real-time 3D reconstruction system.Finally,we developed a 3D reconstruction system,which can measure length automatically and handle massive data of image and point cloud.This paper introduces the principle and method of camera calibration,stereo matching,disparity solving,point cloud fusing,and so on.We use a SIFT-based algorithm in the multi-camera calibration step.It can output internal and external parameters for all cameras in one step.This paper uses a stereo matching algorithm based on MAP.We improved the FAST feature descriptor,combined it with SIFT and color moments.The new combined feature descriptor make the matching process faster and more accurate.The experiment showed that the MAP–based stereo matching algorithm can obtain dense point cloud on the basis of appropriate prior information.We use three methods to make the point cloud fusing more precise.They are external parameters,ICP and manual adjustment.The camera's parameters can change when 3D reconstruction system running in harsh environment.The changes lead to malposed lines and terrible reconstruction result.This paper presents an adaptive line alignment methods based on edge matching,which can resist small-scale changes of parameters and enhance the system's robustness.The proposed system's processing speed is about 2~3 fps on four computers(i7-4790 3.6GHz)and six cameras(5MP).The measurement error is less than 3.5% under automatic mode and 0.14% approximately under manual mode.