A Direct and Robust Method for Ego-Motion Estimation

In general, the motion of the a camera is recovered through dense optical flow or sparse optical flow methods. However, these methods have high computational complexity, and sparse feature points are always usually unreliable in real video sequences. In this paper, we propose an effective and robust ego-motion estimation method based on geometric and photometric constraints, which can directly estimate the parameters of camera motion from video and avoid to calculate optical flow. Additionally, in order to improve the robustness against outliers, we propose a coarse-to-fine iteratively reweighted least squares strategy by introducing error entropy and textural degree to measure the confidence of image patch matching. Experimental results on synthetic and real challenging videos demonstrate that our method can effectively and reliably estimate ego-motion, especially in when there are partial occlusions such as blockage of vehicles and shaded areas.