Real-Time Dense and Accurate Parallel Optical Flow using CUDA

Abstract

A large number of processes in computer vision are based on the image motion measurement, which is the projection of the real displacement on the focal plane. Such a motion is currently approximated by the visual displacement field, called optical flow. Nowadays, a lot of different methods are commonly used to estimate it, but a good trade-off between execution time and accuracy is hard to achieve with standard integrations. This paper tackles the problem by proposing a parallel implementation of the well-known pyramidal algorithm of Lucas & Kanade, in a Graphics Processing Unit (GPU). It is programmed using the Compute Unified Device Architecture from NVIDIA corporation, to compute a dense and accurate velocity field at about 15 Hz with a 640 480 image definition.