Evidential grids information management in dynamic environments

Abstract

An occupancy grid map is a common world representation for mobile robotics navigation. Usually, the information stored in every cell is the probability on the occupancy state. In this paper, an evidential approach based on Dempster-Shafer theory is proposed to process the information in accordance with the least commitment principle. The map grid is updated by a fusion mechanism by using an inverse model of the sensor. We show that the evidential framework offers powerful tools to make a good management of uncertainties especially when the sensory data are poor in terms of information. After having presented the key concepts of evidential grids with respect to probabilistic ones, entropy and specificity metrics are introduced to qualify the degree of information stored in the cells. Some comparisons with the probabilistic approach are given on fusion and decision results using simulation. We also report experimental results to illustrate the performance of a real-time implementation of the method with a 4-layer lidar mounted in the bumper of a car driving in real urban traffic conditions.