A Model for Machine Perception in Natural Environments

Abstract

This paper proposes a new model for machine perception in natural outdoor environments. The goal is to find a method of describing and fusing information obtained from a diverse array of sensors, and to provide a robust interpretation of unstructured outdoor scenes, primarily for the purpose of autonomous navigation. The model has three key components: First,it is based around a high- dimensional sensor-centric description of the environment and deliberately avoids the fragile process of feature extraction from any one sensor. Second, the description is itself embedded in a probabilistic structure allowing Bayes Theorem to be used for temporal fusion of information. Finally, a task-directed process of nonlinear dimensionality reduction, abstraction or compression is used to identify low-dimensional high-contrast features embedded in the high-dimensional sensor space. Such features are the “sensors own view” of what constitutes important information for the task at hand. The paper provides a number of preliminary experimental results of applying this model to combinations of mm-wave radar, laser and night-vision sensors.