Face Tracking and Recognition with the Use of Particle-Filtered Local Features

Abstract

A consistent particle filtering-based framework for the purpose of parallel face tracking and recognition from video sequences is proposed. A novel approach to defining randomized, particle filtering-driven local face features for the purpose of recognition is proposed. The potential of cumulating classification decisions based on the proposed feature set definition is evaluated. By applying cumulation mechanisms to the classification results determined from single frames and with the use of particle-filtered features, good recognition rates are obtained at the minimal computational cost. The proposed framework can operate in real-time on a typical modern PC. Additionally, the application of cumulation mechanisms makes the framework resistant to brief visual distortions, such as occlusions, head rotations or face expressions. A high performance is also obtained on low resolution images (video frames). Since the framework is based on the particle filtering principle, it is easily tunable to various application requirements (security level, hardware constraints).