Error-metrics for Camera Ego-motion Estimation

This paper presents a scheme of camera ego-motion estimation through locating the focus of expansion (FOE). We showed that the bilinear constraint [2] leads to a suboptimal solution of motion parameters in the sense that it does not correspond to maximum likelihood estimate. The contribution of the paper is that we study different error metrics, evaluate the metrics, and propose to use two normalized error metrics under dependent and independent noise model, respectively. They are demonstrated to be optimal in the sense of maximum likelihood. In addition, based on the bilinear nature of the objective functions, we propose to use some specific optimization algorithms to achieve efficient and accurate convergence. Robust estimation problem is also addressed to handle outliers caused by independent motions. Promising results have been obtained in experiments. The estimated motion parameters can be used to detect various independently moving objects on the road.