Dynamic analysis of image motion for vehicle guidance

We outline a selective analysis approach to motion estimation that promises to provide the precision and efficiency required for autonomous vehicle guidance. Efficiency is achieved by implementing computations within a hierarchical (pyramid) structure, and by restricting these computations to selected regions of the scene. These; analysis regions are moved dynamically over the scene as a sequence of focal probes, much as a human driver move his or her eyes and shifts visual attention. Precise motion estimates are obtained by fitting models comprising one or two rigidly moving surfaces to the image data within each focal analysis region. Differential motion within the region separates foreground from background objects, while overall region motion relative to the focus of expansion determines distance from the observer. High level intelligent control directs the focal probes. We show through examples that model-based motion estimation can be used to detect obstacles in the road, and to discriminate such obstacles from road markings. High level intelligent control is described briefly.