Øø Blockinøøóò Òò Ìöö Blockinòò Óó × Òò Ð Øùöö×

Antonio Colmenarez Brendan Frey Thomas S. Huang Adaptive Systems Department Department of Computer S ien e Be kman Institute Philips Resear h University of Waterloo University of Illinois Briar i Manor, NY 10510 Waterloo, Ontario N2L 3G1 Urbana, IL 61801 USA Canada USA Abstra t We des ribe a real-time system for fa e and faial feature dete tion and tra king in ontinuous video. The ore of this system onsists of a set of novel faial feature dete tors based on our previously proposed Information-Based Maximum Dis rimination learning te hnique. These lassi ers are very fast and allow us to implement a fully automati , real-time system for dete tion and tra king multiple fa es. In addition to lo king onto up to four target fa es, this system lo ates and tra ks nine fa ial features as they move under faial expression hanges. 1 Introdu tion In this paper, we present in detail a fully automati , person-independent, real-time system for dete tion and tra king multiple fa es and nine fa ial features. We use Information-Based Maximum Dis rimination lassi ers [1, 2℄ with a novel set of low-level image features to lo ate a urately and eÆ iently nine fa ial features in luding non-rigid points as they move under fa ial expressions. 2 Tra king and Motion Analysis Visual obje t tra king and motion analysis from video are areas of great importan e in omputer vision. The main obje tive of tra king is to roughly predi t and estimate the lo ation of the target obje t in ea h frame of the image sequen e despite hanges in the obje t's pose, size, illumination and appearan e. Motion analysis is on erned with the estimation of the non-rigid motion within the parts of the obje t being tra ked. Figure 1 illustrates a general s heme for obje t tra king and motion analysis. Note that we have highlighted two di erent loops: (i) the tra king loop, whi h exe utes at the frame rate of the input video, and (ii) the initialization loop, whi h exe utes only at the beginning or when the on den e level of the tra king Video Input Model Fitting Tracking Output