Stochastic Classification of Urban Terrain for Smart Wheelchair Navigation

In this paper we investigate the potential for enhancing the perception capabilities of smart wheelchair systems (SWS) in an urban environment through intensity-based terrain classification. A 3D camera system was employed that provided both range and intensity (remission) measurements to objects in the scene. Each sensor measurement was represented by a feature vector consisting of range ρ, bearing angle φ, and intensity γ values. We then constructed a stochastic model by discretizing the feature space into cells, each of which corresponded to a probability mass function (PMF) denoting the probability a point in the feature space belonged to one of four classes: grass, sidewalk, street (blacktop), or unknown. The stochastic model was generated a priori from literally millions of labeled point measurements of the different terrain classes. Bayesian filtering techniques were then used to make a terrain class assignment for ground plane cells in the local map over time. We illustrate the utility of the approach by demonstrating autonomous sidewalk following of the SWS over two hundred of meters solely through terrain class assignment. Extensions such as enhanced ground plane segmentation are also discussed.