Recursive pose dependent AAM: Application to drivers' monitoring

Abstract

A general purpose driver's monitoring system should be able to robustly estimate the main rigid and elastic facial movements through time. Model-based methods have many advantages over feature-based methods for succeeding in this task. This paper presents a system for robust real-time tracking of a set of facial landmarks based on Active Appearance Models. The main differences with other AAM proposals in the literature are twofold. First, a run-time adaptation of the AAM regression matrix using a recursive algorithm to improve convergence precision, second, a multiresolution pose-dependent strategy, PD-AAM, to reduce error in landmarks location for rotated faces and to reduce also computational burden. The proposal is tested in the BUHMAP database, a public database with head and expression movements and in an own set of videos captured in a car. Tests show that the conjunction of these two strategies improve results over a classical AAM tracking system.