Application of Segmented 2-D Probabilistic Occupancy Maps for Robot Sensing and Navigation

The concept of probabilistic occupancy maps was introduced by the end of the 1980s. Over the years, research has focused on the definition of the representation, the data fusion, and the generation of such occupancy models. However, few considerations have been given to processing occupancy maps as textured images to extract meaningful information that is required for robot navigation. This paper investigates the application of modern segmentation techniques over 2-D probabilistic occupancy maps that are encoded as textured images. Enhancements are proposed to a uniformity estimation technique based on local binary pattern and contrast (LBP/C) to achieve the robust segmentation of occupancy maps that typically result from range sensors with limited resolution. The enhanced LBP/C segmentation technique handles occupancy uncertainty and subdivides the space in regions that are characterized by three deterministic occupancy states, which are defined as free, unknown, and occupied. The approach is also extended to increase the number of classification levels, which provides the necessary flexibility to automatically select the regions that are characterized by a given range of occupancy states. The use of these extensions, along with the accuracy of the segmented 2-D occupancy maps, is first experimentally demonstrated on ground-based probabilistic grids for application in mobile robot navigation with collision avoidance. The potential of the proposed approach is also evaluated on aerial and satellite images for which it provides stable results and can find applications for unmanned aerial vehicle navigation.