Sparse Scene Sampling for Robot Vision

T.C. Folsom (USA)


Vision, Robots, Edge finders, Quadrature filters, Interest points, Steerable filters.


Robot vision requires real-time analysis of visual scenes. The visual problem is somewhat simplified because it is ego-centric, multiple frames are available, and the environment may be constrained. Recent work has shown success in using low-level interest points for feature recognition. We present a computationally efficient method for extracting interest points from an overlapping tiling of a visual frame. The method is based on using steerable filters in quadrature pairs. The pair's phase relationship is used to localize edges to subpixel accuracy. This forces a simplified view of the scene since only one feature is permitted in each tile. It also supports curve generation and adaptive allocation of computational resources to the visually richest portions of the scene.

Important Links:

Go Back