Speeding up Dense Optical Flow Estimation with CUDA

Abstract

Optical flow is the perceived movement of a pixel within the video. It is inherently helpful for motion tracking and may also be used as a preprocessing for other computer vision algorithms or machine learning. Algorithmic optical flow estimation is slow and resource-intensive, but real-time performance can be achieved by using GPUs. This paper discusses implementing and optimizing a pyramidal Lucas-Kanade optical flow algorithm in CUDA.