Sequential Learning for Robot Vision Terrain Classification

Terrains have widely varying visual appearance depending on the type of foliage, season, current weather conditions, recent precipitation, time of day, relative direction of lighting, presence of man-made structures and artifacts, landscaping, etc. It is difficult, if not impossible, to specify in advance the appearance of the different terrains that will be encountered while operating a robot in urban or rural environments. Yet people, having accumulated wide-ranging experience, have little trouble recognizing familiar terrain types and learning to recognize new, previously unfamiliar, terrains. Robots typically accumulate experience in "chunks" and do not have the luxury of years of training. This paper presents recent results in sequential learning methods applied to robot terrain recognition. In this paper we explore different sequential learning problem formulations and alternative machine learning algorithms. The investigations are based on the same data set. We report on the initial development of an incremental fuzzy c-means clustering algorithm capable of learning new information. We report on an approach to convert regression tree modeling, normally a batch learning method, to batch-incremental learning. We investigate issues in formulating the sequential learning problem and the performance of these algorithms. We also compare performance to four incremental learning classifiers. All investigations were conducted using the same set of image features, extracted from on-board video from a small robot traversing different terrains.