Visual Odometer Based On Binocular Vision

With the continuous development of computing ability,computer vision related technology has come into use.Currently,the research of vision based environment sensing technology is hot,such as Visual Odometry,which is also the key technique in unmanned vehicle,AR and other applications.This article is devoted into the details of Visual Odometry(VO).With information acquired via a binocular camera,and on the framework of ORB-SLAM2,we got some improvements on accuracy and made it fit the structured indoor environment.The main work of this article is as follows:(1)We migrated the wide-spread ORB-SLAM2 onto the mobile robot hardware platform,which involves the work of rewriting the image input interface,the calibration of the point-grey BB2 binocular camera and image rectification.(2)In an insightful view of the failure of conventional VO in a structured environment,for example in buildings,the SURF features were modified,thus we got the mixed Incomplete SURF(I-SURF)and SURF feature.A series of experiments were carried out under the self-built structured scenes to prove its effectiveness.What's more,We also compared it with other methods on the classic KITTI dataset.(3)With regard to the "drop" often occurred in VO,we focused on the relocating process,and enumerated two repositioning schemes that followed by several experiments to make a comparison.The innovation of this paper is to modify the general VO scheme to adapt to the indoor structural environment,particularly,the above(2)(3)points.After revision,the proposed method got through several scenes in the self-built real environmental data set where libVIS02,ORB-SLAM2,LSD-SLAM would fail.Besides,the relocating accuracy acquired a rise of about 20%higher than that of the above schemes.The proposed method extends the application environment of VO.