The goal of our research is to design perception devices dedicated to driving safety improvement such as : Collision Warning and Avoidance systems, Emergency braking… This device is supposed to be a piece of new driving assistance systems aiming at increasing the safety on road. Many safety systems which are emerging nowadays in our vehicles use a “distance to obstacle” information obtained thanks to telemeters such as radars, laser scanners (or lidars), ultrasounds… At least, this kind of systems achieve a great precision in locating objects relative to the sensors, but they are not able to provide their localisation compared to the road or the lane. This is the reason why some systems are involving passive sensors like video integrated inside vehicles. A device founded partially on computer vision allowed to compensate this lack. But the localisation computed by vision needs to be analysed in term of precision. This paper attempts to explore differences in localisation accuracy between systems involving only one camera (monocular vision) and systems involving two cameras (stereovision). A complete study of the errors commit on depth reconstruction is shown.
Keywords:
monocular vision; stereovision; active safety; depth and distance reconstruction; obstacle detection; retroprojection