Rotation and translation estimation for a wide baseline fisheye-stereo at crossroads based on traffic flow analysis

Many vision-based systems have been proposed for intersection monitoring, but few of them use wide angle cameras. In this paper, we present a fisheye-stereo monitoring system that will be installed especially in rural areas. Our goal is to estimate the extrinsic calibration decoupled into a rotation followed by a translation. In the context of rural intersections, methods that purely rely on the geometry of the scene are difficult to apply. In fact, the presence of vegetation and the large baseline between the cameras make the task particularly challenging. We present instead an approach based on vehicles' motion analysis on the main road, along with the generation of virtual lines joining the front wheels. We adopt the spherical camera model and use the assumption of planar motion. The approach is validated on a test drive field with a baseline larger than 13 m. We get errors below 3 degrees for the rotation and less than 0.5 percent for the translation at scale. Our method allows a reliable initialization for 3D intersection monitoring.