Research Of Monocular Visual Odometry Based On Semi-Direct Method

The technology of Visual Odometry(VO)and Visual Simultaneous Localization and Mapping(VSLAM)use a serial of image sequence to locate the 6-DOF camera pose in space,which becomes increasingly popular in the applications of indoor robots,drones,Augmented Reality and Virtual Reality.However,onboard computing capability for robots is limited,so the accuracy and real-time processing is essential for VO as a basic function module.There are two different error models for pose solving in Visual Odometry including reprojection error model the photometric error model.The method which uses those two models is so-called feature-baseds method and direct methods.The feature-based methods extract features and compute descriptors for each image,and match those features by descriptors.However,the direct methods estimate motion directly from the error computed by the brightness of image pixels which is no need of feature extraction for each image and without knowing feature's correspondence in advance.Giving consideration to the speed and precision,this thesis utilizes a semi-direct method,using the direct-based method to estimate initial frame-to-frame motion and features matching,while using bundle adjustment to refine camera's pose basing reprojection error subsequently.This thesis uses inverse depth filter to recover features' 3D location.FAST corners are extracted in each keyframe and created as seeds,where inverse depth is modeled by Gauss+Uniform mixture distribution.Those seeds are matched by epipolar line and optical flow.The inverse depth of seed is updated by measure from triangulation.The experiments show that the proposed seed tracking strategy is effective and the inverse depth filter model can decrease the outlier ratio.For the consideration of improving precision.every time a key frame created,the map points will be maintained and the bundle adjustment will be applied for the local map.This thesis uses a sin gle local mapping thread.in which adds more measurements for keyframes by feature alignment,fusion identical map points and reject outliers after bundle adjustment.There are 10 keyframes selected which shares most observations with new keyframe and of which the map points observed are used for local bundle adjustment.In experiments,the algorithm proposed by this thesis compared with the state-ofarts VO/VSLAM both in precision and running time.Extensive experiments indicate that the semi-direct visual odometry designed in this thesis achieve outperform in speed with more than 50Hz's output rate for camera pose,and its precision is close to ORB-SLAM2 which has highest precision.