Learning-based architecture for robust recognition of variable texture to navigate in natural terrain

Abstract

Since natural terrain consists of textured objects that can be perceived under different external conditions, the author applies machine learning methodology to support the recognition of variable texture. He presents the results of first introductory experiments and the development of a new system architecture incorporating learning tools; i.e. conceptual clustering, learning from examples, and learning flexible concept. He then describes the designing methodology and system architecture of three functional levels typical for the large-scale control systems; i.e. self-tuning to a given content of texture image in order to extract most sensitive features and to group them into patterns, learning a concept of new texture, and control of system adaptation (guided by vision goal, feedback verification of created hypotheses, and a plan of the environment content). He also discusses the requirements for learning tools that are used to build such adaptive vision systems and presents their further development.<>