Particle filters for multi-face detection and tracking with automatic clustering

Abstract

We develop a systems that detects and track individuals and groups of people in real-time, designed as a first screening of the iris-based access control. We use particle filtering in the conditional density propagation framework of hard and Blake, and base the face detection on the skin color in HSV colorspace. Particle assignment to faces is done automatically the use of X-Means algorithm. The X-means algorithm is modified to (1) avoid the local extrema (2) to deal better with odd number of clusters. The modified algorithm is fast enough to work in real time and outperforms its original version.