Flat world homography for non-flat world on-road obstacle detection

The goal of our research is to design perception devices dedicated to driving safety improvement such as: collision warning and avoidance systems, emergency braking, etc. The devices are supposed to be a part of new driving assistance systems aimed at increasing the safety on the road. This paper explores a new computer vision method for detecting obstacles above the road's surface. This new approach must be considered in an exploratory phase. We try to show the potentiality of this approach when knowing that many problems must be solved at the moment. It runs without assuming small rotations and translations between images. The algorithm has a homographic formulation for the flow of pixels between two consecutive images. We warp the image at time (t-1) to match it on the image at current time (t) according to the movement of road's surface. The innovation of the system is that we do a cooperation between inertial sensors and a video camera to achieve the warping and the detection. The movement in images is compensated, so that we deal with a "pseudo-stereo case" and apply a temporal "v-disparity" post-process.