Biological vision models for sensor fusion

A general-purpose machine vision system for sensor fusion is described. The approach taken is to model biological vision with neural networks and traditional processing. The architecture has two channels. A locating channel searches for objects in the field-of-view. A classifying channel learns and recognizes the objects. Learning is by example. The architecture can fuse disparate sensory inputs, including one- and two-dimensional data, at the feature level. Fusion was tested by computer simulation. The sensor inputs were two-dimensional resolved images of vehicles from coregistered video intensity and laser radar range sensors. The study produced probability transition tables for three vehicles, with and without sensor fusion. Results suggest that fusing sensors in this way gives a robust, practical, and high-performance machine vision system.<>