A full-pixel optical flow system using a GPU-based high-frame-rate vision

Qingyi Gu, Naoki Nakamura, Tadayoshi Aoyama, Takeshi Takaki, Idaku Ishii

Abstract: This paper reports real-time full-pixel optical flow system that can simultaneously estimate pixelwise motion distributions of 512×512 images by accelerating the Lucas-Kanade method on a GPU-based high-speed vision system. In our optical flow system, the measurable dynamic range of the estimate optical flows is remarkably expanded by introducing a novel algorithm that can optimally select space intervals in optical flow estimation according to the estimated flow speed. Several experimental results for high-speed objects are shown to verify its effectiveness in high-frame-rate and real-time optical flow estimation.