視覚神経の時空間応答特性に学んだリアルタイム両眼視差検出システム

抄録

We have developed a real-time binocular disparity detection system by combining the spatiotemporal properties of the visual nervous system and the binocular energy model. This system consists of two image sensors, three field-programmable gate arrays (FPGAs), a USB controller, and a personal computer (PC). The spatio-temporal properties implement motion direction selectivity into the binocular energy model, and can reduce the number of binocular matching candidates. We evaluated the system by presenting an object rotating at a constant angular velocity on a motor-controlled turntable to the system. While the original binocular energy model estimated a wrong disparity due to mismatching when the rotating object moved near a background object that has similar spatial features, the binocular disparity detection system estimated the correct disparity in the same situation because the spatio-temporal filter prevented the mismatching.