Approximated Relative Pose Solvers For Efficient Camera Motion Estimation

We propose simple and efficient methods for estimating the camera motion between two images when this motion is small. While current solutions are still either slow, or unstable in case of small translation, we show how to considerably speed up a recent stable but slow method. The reasons for this speed-up are twofold. First, by approximating the rotation matrix to first order, we obtain a smaller polynomial system to be solved. Second, because of the small rotation assumption, we can use linearization and truncation of higher-order terms to quickly obtain a single solution. Our experiments show that our approach is both stable and fast on challenging test sequences from vehicle-mounted cameras.