Motion estimation with a mobile camera for traffic applications

This paper describes a digital rear-view mirror that assists the driver to analyze situations on the road behind him. From the observations of a single camera we want to derive the real-world motion of the vehicles behind our car. These observations are influenced by the egomotion of the camera. In our approach, we use a robust algorithm to estimate the mean viewing direction of the camera. Provided with this information we correct the observations on the fly to simulate the situation where the camera is looking straight backwards. This makes further processing algorithms less complex. We explain how to estimate real-world motion parameters of interest, like the time-to-contact and lane-shifts. Results based on practical data are presented to indicate the practical validity of our approach. The experiments show how the viewing direction is corrected in an accurate and robust way. The estimation procedure for the time-to-contact is improved by the correction algorithm for the viewing direction, but still requires some refinement.