GPU-Accelerated High-Speed Eye Pupil Tracking System

Pupil tracking under infrared illumination is an important tool for many researchers in physiological visual optics and ophthalmology. It is also a relevant topic for gaze tracking which is used in psychological and medical research, marketing, human-computer interaction, virtual reality and other areas. A typical setup can be either a low-cost webcam with some infrared LEDs or glasses with mounted cameras and infrared illumination. In this work, we evaluate and parallelize several pupil tracking algorithms with the aim of estimating the pupil´s position and size with high accuracy in order to develop a high-speed pupil tracking system. To achieve high processing speed the original non-parallel algorithms have been parallelized by using CUDA and OpenMP. Graphics cards are designed to process images at very high frequencies and resolutions, and CUDA enables them to be used for general purpose computing. Our experimental results show that pupil tracking can be efficiently performed at high speeds with high-resolution images (up to 530 Hz with images of 1280x1024 pixels) using a state-of-the-art GP-GPU.