Advanced Model Based Vision System for Intelligent Transportation System

An advanced model based vision system for intelligent transportation systems

Personnel

M. Kais (C. Laugier, M. Parent, I. Masaki and B.K.P. Horn)

Sponsorship

Intelligent Transportation Research Center, Lounsbery Foundation, INRIA

In order to offer better safety and increase the capacity of the roads, new concepts of mobility are being developed. These advanced transportation systems are based on autonomous driving and platooning. Platooning applications, consist of creating a platoon of electronically coupled vehicles with a very small headway. In a platoon, the first vehicle is manually or automatically driven and the others follow. Another application is driverless fully autonomous electric vehicles called cybercars (see fig1) for transportation in urban environment.

These applications require sensors for the guidance and obstacle detection tasks. The goal of the vision system is to build an accurate representation of the environment in front of the vehicle to be used by the planning layer. Stereo vision offer a low cost and easy way to get some range information about the environment. Traditional vision systems are fine for the Highway environments, because they are well defined (size, marker) however urban environments present a challenging task due to the complexity of the scenes. For instance, in a single frame, the road and lane boundaries can take on several primitives, such as a curb, a white lane marker, or a line of cars parked on the side of the road.

Our approach consists in fusing information from the cameras with some a priori knowledge stored in a database. This a priori knowledge consist of a global model i.e. an environment model (how the roads are linked), a geometric model of the road and lane boundaries (shape) and a model of the road and lane boundaries feature (lane marker, curb, planes). This global model is acquired during a learning phase and linked to an existing geographic database using a Geographic Information System. Once the learning phase is completed, the estimate of the position of the vehicle and the a priori knowledge are used to enable some specialized detectors in specific Regions Of Interest (see fig 2) in order to extract from the images relevant information that is used for the 3D reconstruction.

Fig1 : A cybercar

Fig2: Results from the lane marker detector