Detection and tracking faces in unconstrained color video streams

Abstract

This paper proposes a method combining local SVM classifiers and a Kalman filter to track faces in color video sequences, which is referred to as the Dynamic Local Support Vector Tracker (DLSVT). The adjacent locations of the target point are predicted in a search window, reducing the number of image regions that are candidates to be faces. Thus, the method can predict the object motion more accurately. The architecture presented good results for both indoor and outdoor unconstrained videos, considering multi-view scenes containing partial occlusion and bad illumination. Moreover, the reduction of the image area in which the face is searched for results in a method that is faster, besides being precise. © 2011 Springer-Verlag.