Low-Cost, Real-Time Face Detection, Tracking and Recognition for Human-Robot Interactions

Book excerpt: This dissertation presents various vision-based algorithms for human-robot interactive applications, such as sociable robots. Our vision-based methodologies include accelerated AdaBoost classifier based face detection, self-learning face tracking and adptive PCA-based facial recognition. By using a resizing technique and skin tone filter, we only apply the AdaBoost classifier to a small region and thus, compared to applying it to the whole image, require much less processing time. In order to track a detected face precisely and efficiently while also recognizing the detected face, a hybrid face tracking approach is applied based on an adaptive skin color mode and a potential window. A novel adaptive face recognition method is implemented by automatically upgrading the set of sample faces of a 0́known0́+ person and collecting information of 0́unknown0́+ people for enhanced recognition performance. Some additional effort has been made on speech recognition system, voice system and behavior system. These algorithms are well suited for embedded systems because of their cost and time efficiency and little pre-training required for reliable performance. All of the above algorithms have been tested on a sociable robot named 0́−Philos0́+ developed in the Distributed Intelligence and Robotics Laboratory at Case Western Reserve University.