A system to detect the car driver's hypovigilance states, based on the measurement of blink parameters by digital image processing and neural networks is being developed. The images are captured by a camera which is not attached to the driver. To get blink parameters, the eyes have to be classified into two states: open or closed. To reach this goal, the eyes must be automatically located in the images. This paper concentrates on this location problem. In our system, the location of the eyes is based on the preliminary location of the head. We use a neural network to carry out a Principal Component Analysis (PCA), a useful technique to search for a given pattern in a signal or an image. PCA is applied to small images of heads and eyes (of size 32x32 pixels), obtained from 256x256 source images captured by a CCD camera. With neural networks, PCA is made in an intrinsic parallel way and by means of a learning phase. The images of heads and eyes are pre-processed by convolution with a Laplacian-of-Gaussian operator before being learned. After learning, another network of processing units, the search network, is used to search for the head and then the eyes in test-images. Simulations show that eye location is accurate in images showing people not presented to the learning network during the learning phase. (A)
Abstract