Robot Visor Research: Past and Future Roles

This paper describes what I have done and learned in robot-vision research, and speculates what is necessary and promising in the future. In early artificial intelligence research, it was proved in the context of recognition of polyhedra that a hierarchical method in which a higher process assumes an ideal output of lower processes has limitations. A similar approach has been effective for vision processes that use the result of color segmentation, stereo vision, optical flow, or correspondence between multiple images. It is often more effective for recognition of difficult scenes to increase available input data than to try to make a clever procedure. There was an attempt to approach more flexible vision systems like human vision, which include reliable stereo vision and integration of multiple visual cues. Now a tightly coupled perception-action paradigm is an important issue, where significant research themes are person tracking and recognition, flexible real-time vision processors, and planning of perception-action considering the planning cost.