Asahi Kasei - Driver drowsiness recognition

The goal of this project was to recognize if a vehicle driver is sleepy by (i) finding the driver’s face and eyes; and (ii) classifying eye blinks to estimate drowsiness.

Face detection is performed from infrared images using a Support Vector Machine (SVM). The detector is trained with frontal face images and rotated examples of up to 90 degrees to either side. To achieve real-time performance, the number of Support Vectors is reduced using the Reduced Set Method. Furthermore, PCA is employed to reduce feature dimension. Subsequently, the position of the eyes are found with a similar method.

Once the position of the eyes are known, eye blinks classification is performed. We distinguish three different types of blinks: 'normal', 'long' and 'short'. The latter two types are more often observed in sleepy subjects and hence are used to estimate the drowsiness of a driver. For each blink type, a Hidden Markov Model (HMM) is learned and the classification accuracy of different features is evaluated.

Example videos: faceDetection.avi normalEyeBlink.avi (xVid Codec)