Feature-based perception for autonomous unmanned navigation

This paper presents a fuzzy-based approach to robot perception based on feature maps generated using a commercial off-the-shelf (COTS) stereovision system, intended for the guidance of autonomous vehicles in urban and outdoor environments. Feature maps integrate colour with geometrical information onto an occupancy grid, for representing useful features extracted from the perceived environment. These include obstacles, traversable regions, and objects-of-interest detected from disparity images. A likelihood measure of the presence of obstacles within the perceived environment is introduced by comparing the displacement of each grid cell relative to a dynamically-estimated virtual ground plane. With a priori knowledge of the colour compositions of traversable (and non-traversable) regions, a fuzzy colour segmentation algorithm is used to extract such regions, thereby enhancing the autonomous vehicle's knowledge of the environment. This method is used for computing the degree of safety in traversing through the local environment. The advantages of this approach in computational speed, ease of implementation and obstacle detection robustness based on multiple features are supported by promising experimental results obtained on both live and recorded images.