Pavement Boundary Detection Via Circular Shape Models

Automated detection of pavement boundaries is an important enabling technology in a number of intelligent vehicle applications. Many state of art systems for detecting and tracking pavement boundaries use a priori shape models to describe the appearance of these boundaries. Several types of shape models have been employed, and the model choice is usually made from the standpoint of accommodating all possible variations (in width, orientation, curvature, tilt, etc.) of the pavement boundaries relative to the host vehicle. Polynomial (quadratic or cubic) shape models are the ones of choice. This paper describes a circular shape model for detecting pavement boundaries. Indeed the polynomial shape models are intended as an approximation to the circular model, but the circular model itself has never been used before. This paper shows that the circular shape models enjoy several critical advantages over the polynomial models without any additional increase in model complexity (i.e., number of model parameters): the model parameters are all of the same units, even a small change to any one parameter results in a uniformly different shape appearance, and as a result the associated shape matching problem (the one of matching various pavement shapes to the observed image) is considerably better conditioned than the corresponding problem with polynomial shape models. Our application domain is one of road/pavement boundary estimation based on image data from a high-resolution multibeam 77GHz millimeter-wave radar. A successful solution to this problem will impact a number of driver assistance systems, such as road departure warning, forward collision warning, etc. Research has been sponsored by the National Automotive Center at U.S. Army TACOM, and the Defense Advanced Projects Agency. Proceedings of the 2000 Intelligent Vehicles Conference, The Ritz-Carlton Hotel, Dearborn, MI, USA, October 3-5, 2000.