Appearance-based monocular visual odometry for ground vehicles

This paper presents a method for computing the visual odometry of ground vehicles. By mounting a downward-looking camera on the vehicle, it can recover the motion of the vehicle only by using the video input. Different from other feature based algorithms, the key technique deployed in this system is an appearance-based approach - template matching. A rotated template matching approach is used for motion estimation. For each two consecutive images captured from the camera, an image patch is first selected from the first image, and then the image patch is rotated by every angle within a predefined range and matched against the second image. The best result from the matching indicates both the position translation and heading changes. For performance evaluation, this algorithm is tested on the indoor granite surface. In the end, we give a comparison of the result between wheel odometry and visual odometry.