A Real-Time Demonstrator for Video-Based Recognition of Dynamic Head Gestures, Using Discrete Hidden Markov Models

This work describes a powerful demonstrator of a videobased approach for detecting and classifying dynamic head gestures. The head of the user is localized via a combination of colorand shape-based segmentation. For a continuous feature extraction, we use a template matching of the nose bridge in combination with selected features derived from the optical flow. The core classification unit consists of discrete Hidden Markov Models (DHMMs). We extensively tested the system in two different domains (desktop VirtualReality and automotive environment). In the current state of development, six different gestures can be classified with an overall recognition rate of 97.3% in the VR, and 95.5% in the automotive environment, respectively. The approach works absolutely independent from the image background and additional gesture types can easily be integrated.