Face Tracking by Means of Continuous Detection

Abstract

The main contribution of this work is a new view to face-tracking namely to associate the independent detection results obtained by applying a real-time detector to each frame of a sequence. This differs fundamentally from traditional tracking which is mostly understood as finding the location of a target object given its previous position and a correspondence finding algorithm either with or without a prior object model. This traditional notion has two major problems namely the initialization problem and the lost-track problem. We show that the advent of new rapid detection algorithms may change the need for traditional tracking. Furthermore the mentioned problems have a natural solution within the presented tracking by continuous detection approach. The only assumption on the object to track is it's maximal speed in the image plane, which can be set very generously. From this assumption we derive three conditions for a valid state sequence in time. To estimate the optimal state of a tracked face from the detection results a Kalman filter is used. This leads to an instant smoothing of the face trajectory. It can be shown experimentally that smoothing the face trajectories leads to a significant reduction of false detections compared to the static detector without the presented tracking extension. We further show how to exploit the highly redundant information in a natural video sequence to speed-up the execution of the static detector by a temporal scanning procedure which we call slicing. A demo program showing the outcomes of our work can be found in the internet under http://www.iis.fraunhofer.de/bv/biometrie/ for download.