Three dimensional image stiching method ofrobot binocular vision based on double registration
                                                                   AI Qinglin, LIU Sai, SHEN Zhihui
                           (College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310014, China)

Abstract: Aiming at the problem that the vision field of the robot image is not wide enough to obtain comprehensive obstruction information, a 3D reconstruction and variable visual splicing method based on contour recognition was researched. The edge points were extracted from the two images taken by the binocular stereoscopic vision system and subjected to boundary tracking. Then the corresponding pixel points on the two images were matched to reconstruct 3D contour of the obstacle based on the window gray scale matching method. The 3D contours were optimized according to the continuity of the target object contour. On the basis of this, a new method of variable view threedimensional splicing based on double registration was proposed. In order to obtain obstacle information in the large field of view, two 3D point clouds were converted to the same coordinate system for accurate match and image stitching, according to the improved ICP algorithm. The results show that the 3D model contour spliced by this method can obtain the highprecision and large field image by changing the angle of view, which can meet the requirements of robot obstacle detection and path planning.

Key words: binocular stereoscopic vision; obstacle detection; 3D reconstruction; double registration method; 3D image stiching