Obstacle detection by recognizing binary expansion patterns

This paper describes a technique for obstacle detection, based on the expansion of the image-plane projection of a textured object, as its distance from the sensor decreases. Information is conveyed by vectors whose components represent first-order temporal and spatial derivatives of the image intensity, which are related to the time to collision through the local divergence. Such vectors may be characterized as patterns corresponding to “safe” or “dangerous” situations. The authors show that the essential information is conveyed by single-bit vector components, representing the signs of the relevant derivatives. The authors use two previously developed, high capacity classifiers, employing neural learning techniques, to recognize the imminence of collision from such patterns