Dealing with the structured scene in visual odometry(VO): Incomplete SURF

摘要

In this paper we study the Visual Odometry (VO) in the structured environment. Harris, SURF, ORB, such feature extraction methods are often employed by current VO methods. They are good at corner detection, but, tend to fail in an environment full of structure lines. Here we focus on just one direction of the gradient, and modify the SURF method to make it fit in the specified scene, namely, the Incomplete SURF method. As it is supposed to run in a VO framework, the calculation of the transition matrix between frames has to be adjusted accordingly. We tried the proposed combined SURF and Incomplete SURF method in our self-established dataset (seldom corners, a lot of structure lines), and succeeded in estimating the transition matrix where pure SURF failed.