High Speed Eye Tracking Using The Vision Chip

This work demonstrates the implementation of a high speed eye tracking system capable of tracking saccadic eye movement. Setting out to capture saccades with velocities of up to 700/sec and durations as short as 20ms, the current system provides a frame rate of 100Hz. To achieve this frame rate, the system utilizes low cost, small size tracking hardware: the Vision Chip. The system tracks eye movements based on the pupil position. Segmentation to isolate the pupil is mainly performed by binarization, and the position is determined based on moments of the pupil area within the binary image. The main problem solved in this work, is to avoid cluttering of the binary image due to binarization artifacts, and to detect distortions of the pupil that affect the accuracy of the positional information. A natural cause for distortions is the occurrence of eye blinks, others are caused by the variation of the appropriate binarization threshold for different eye positions. To deal with these problems in realtime, the tracking is preceded by in depth analysis of grayscale images of the eye, to locate the pupil and to determine a pupil related feature vector, including a binarization threshold. This information is acquired for several eye positions, where each threshold is ideal for the position, i.e., few binarization artifacts appear. Using these thresholds, the binarization threshold can be dynamically adjusted during tracking to reduce the amount of clutter. To detect distortions, the system also computes a vector of pupil related features at runtime. Using the precomputed features as constraints, distortions can be detected by comparison. This is achieved with the required frame rate, by taking advantage of special Vision Chip functions that support moment calculations. The thesis demonstrates the calculation of moments up to the second order and gives area and eccentricity as examples for features, the eye position is computed as the pupil centroid. These positions are available in realtime or can be stored for off-line analysis. They are provided as 2D positional information that can optionally be mapped into screen coordinates to acquire Point Of Regard information. The positional resolution has been estimated to be between <0.14 and <2.3 over a range of 45.