Resumo

Drowsiness is a major cause of car accidents and consequently a source of life and property risk. We propose the development of a drowsiness detection system based on a web cam, simple image processing and machine learning techniques. The acquired image is converted from RGB to intensity gray-scale, normalized for intensity and the face is detected using available Matlab tools. Next, the region of the left or right eye is estimated and the pupil location is detected using a gradient contrast technique. The eye open/closed state is estimated using function fitting to the vertical intensity line that crosses the pupil position. The function parameters and human observer ground-truth are used to train a two-class pseudo-Fisher classifier. Preliminary qualitative results, evaluated by human observers, show that the method is suitable to estimate driver’s blinking in optimal daylight conditions. In the next phase an infrared light at night time conditions will be used to generate long image sequences that will include low and high drowsiness states. Those videos will be quantified for blinking by human observers and extensive tests of the proposed system will be used to accurately indicate the error rate in realistic driving conditions. Apoio: LTCM had a PDPD scholarship from Pró-reitoria de Pesquisa - UFABC.