Efficient visual odometry estimation using stereo camera

In this paper we present a visual odometry system that estimate the motion of a stereo camera from consecutive stereo image pairs. No other sensors or prior knowledge of the scene is needed. The point features are extracted and matched between image pairs. Given feature correspondences, the camera motion is estimated by minimizing the reporjection error of successive images. The main contribution of this paper is that we design a customized feature descriptor that can make feature matching process extremely fast while preserving the matching reliability. Hence, we can include as many features as possible into optimization process which leads to improvement of motion estimation accuracy. Real-world experiments are done to demonstrate the ability and efficiency of the system.