Feature Matching Combining Variable Velocity Model with Reverse Optical Flow

The ORB-SLAM2 based on the constant velocity model is difficult to determine search window of reprojection map points when objects are in variable motion, which leads a false matching, with an inaccurate pose estimation or failed tracking. To address challenge above, new method feature point matching proposed this paper, combines reverse optical flow method. First, extended model, and expanded used provide initial pixel shifting for Then range accurately determined according results method, thereby improving accuracy reliability strengthened interframe tracking effects. Finally, we tested TUM data set RGB-D camera. Experimental show that can reduce probability failure improve localization SLAM (Simultaneous Localization Mapping) systems. Compared traditional ORB-SLAM2, test error each sequence significantly reduced, root mean square only 63.8% original system under optimal condition.