RETINA FOR REAL-TIME COMPUTATION OF IMAGE FLOW

WE PROPOSE TO DEVELOP AN ADVANCED OPTICAL-FLOW MOTION ESTIMATION AND SEGMENTATION ALGORITHM FOR USE WITH MOBILE ROBOTS. BASED UPON OUR EARLIER WORK ON OPTICAL-FLOW IMAGE ANALYSIS, AN ADVANCED OPTICAL-FLOW SEGMENTATION ALGORITHM, USING DISCONTINUITY INFORMATION IN BOTH VELOCITY AND INTENSITY DOMAINS, PERMITS THE SEGMENTATION OF VIDEO SEQUENCES ALONG NATURALLY OCCURING MOTION BOUNDARIES. ESTIMATING THE VELOCITIES OF BOUNDARIES MAY PERMIT MORE ACCURATE MEASUREMENT OF THE FRAME-TO-FRAME COHERENT MOTION IN VIDEO SEQUENCES THAN IS POSSIBLE WITH CURRENT BRUTE-FORCE, BLOCK-BASED ESTIMATION ALGORITHMS. ACCURATE OBJECT BOUNDARY VELOCITIES ARE ESSENTIAL IN ESTIMATING RANGE, DEPTH, AND TIME-TO-CONTACT OF OBJECTS IN VIDEO SEQUENCES. THE INHERENT PARALLEL NATURE OF THE OPTICAL-FLOW CALCULATIONS ALLOWS FOR THE IMPLEMENTATION OF THE ALGORITHM ONTO A LOW-COST ANALOG ASIC. A SMALL CHIP WITH CUCH FUNCTIONALITY WOULD LOWER THE COST OF UNMANNED VEHICLES THAT USE MOTION INFORMATION TO NAVIGATE. DURING PHASE I, WE WILL INVESTIGATE STRATEGIES FOR OPTIC FLOW MEASUREMENT. VARIOUS ALGORITHMS WILL BE STUDIED FROM THE VIEWPOINT OF CORRECTNESS AND SILICON SUITABILITY. WE WILL DESIGN AND FABRICATE A CHIP WITH AN ANALOG MOTION-SENSITIVE PIXEL. WE WILL CHOOSE THE BEST ALGORITHM FOR FURTHER DEVELOPMENT INTO A CHIP WITH FULL-FRAME MOTION ESTIMATION ABILITY IN PHASE II.