In this chapter, the authors elaborate on the facial image segmentation and the detection of eyes and lips using two neural networks. The first neural network is applied to segment skin-colors and the second to detect facial features. As for input vectors, for the second network the authors apply speed-up robust features (SURF) that are not subject to scale and brightness variations. The authors carried out the detection of eyes and lips on two well-known facial feature databases, Caltech. and PICS. Caltech gave a success rate of 92.4% and 92.2% for left and right eyes and 85% for lips, whereas the PCIS database gave 96.9% and 95.3% for left and right eyes and 97.3% for lips. Using videos captured in real environment, among all videos, the authors achieved an average detection rate of 94.7% for the right eye and 95.5% for the left eye with a 86.9% rate for the lips
TopIntroduction
A number of approaches have been proposed to detect facial features. Among all facial features, eyes have become a maverick for their variety of applications and implications. In the case of car driver eye detection, color cameras and cameras with IR illumination have been widely used as a video capturing device. The use of cameras with IR illumination significantly simplifies the problem of eye detection. When an infrared led is located at the camera axis, the eye irises appear as two bright spots caused by the reflection of the blood-rich retina. Thus, IR illumination based approaches gained high popularity in eye detection (Zhao & Grigat, 2006) and consequently driver attention monitoring tasks. In (Ji & Yang, 2002) a system for monitoring driver vigilance is proposed. The main idea is to place two cameras at different angles. The camera axis coincides with two coplanar concentric rings where along their circumferences a number of IR LEDs are evenly and symmetrically distributed. One camera has a wide range view for head tracking, and the other has a narrow view for eye detection. A similar hardware setup was applied by Batista et al. (Batista, 2005), the difference from (Ji & Yang, 2002) consists in face and eye detection algorithms. Hammoud et al. (Hammoud, Witt, Dufour, Wilhelm, & Newman, 2008) proposed a complete driver drowsiness detection system that detects irises in the near infrared spectrum. Although IR based approaches perform reasonably at night time, it was noted (Hartley, Horberry, & Mabbott, 2000) that those methods often malfunctioned during daytime under the presence of sunlight. Moreover, when eyes are closed the reflection in IR range disappears, making eye detection a difficult task. Another disadvantage of IR based approaches is the necessity of installing an IR LEDs setup.
In comparison with IR cameras, CMOS and CCD cameras are passive, meaning there is no IR radiation. The effect of long term IR radiation should be studied to guarantee that there is no danger to eye health (Pitts, Cullen, & Dayhew-Barker, 1980). CMOS cameras are relatively inexpensive and ergonomic. Furthermore, according to (Hartley, et al., 2000) 52% of drivers nodded off while driving between the times of 6:00 a.m. and 9:00 p.m. comprising of the majority of bright daylight on any given day. As a consequence, using IR cameras during those hours is impractical since IR cameras inefficacious under direct sunlight.