Robust ground plane estimation based on particle filters

This paper describes a particle filter based approach for estimating the ground plane from an image sequence. Based on a Bayesian framework, the particle filter provides a robust estimation of the plane parameters, since it can handle non-linearities, while allowing a high flexibility for integrating new cues into the system. Furthermore, the different modes of the resulting probability density function are segmented by means of a mean-shift algorithm, resulting in better localization of the estimate with the highest posterior probability. Our method has been tested on both synthetic and real world scenarios and has shown to be robust to missing and unstable measurements. On synthetic data of representative runs the angular error is well within 0.5deg with a standard deviation of less than 0.3deg.