Image processing-based eyelid state detection

Abstract

This paper presents a new image-processing algorithm for tracking the eyelid states in low-resolution images. The technique is useful in biomedical areas including contactless interfacing, aiding tool in diagnosis of neurological conditions and patient monitoring. The algorithm recognizes frontal faces inside the image using a Viola and Jones fast detection algorithm and identifies the regions of interest (ROI) around both eyes. The scale and rotation-invariant projection is computed and used in binary threshold-based classifier to determine the eyelid in open or closed state. Receiver Operating Characteristics (ROC) curve analysis is applied to evaluate the approach and to define the optimal threshold values. A preliminary study with 18728 frames from 15 subjects was carried out and the distance and subject-dependent errors are discussed in terms of probabilities. The single frame equal error rate ranges from 6.0 to 12.4%. These errors can be reduced by using multi-frame logic and controlling the illumination and face positioning. The developed technique was applied with encouraging results in several rehabilitation systems including: (a) a developed Speech Synthesizer by Eye Clicks (SPEK) used by a physically disabled person whose movements are almost only restricted to the eye blinking to communicate with family members and visitors, (b) a developed Face Mouse System and Mouse and Keyboard Emulation System used to enable severely disabled people to explore any application in a computer. Increased independence level, access to information, and social interaction were observed for the systems users.