Robust video-based recognition of dynamic head gestures in various domains - comparing a rule-based and a stochastic approach

This work describes two video-based approaches for detecting and classifying dynamic head-gestures. We compare a simple, fast, and efficient rule-based algorithm with a powerful, robust, and flexible stochastic implementation. In both realizations, the head is localized via a combination of color-And shape-based segmentation. For a continuous feature extraction, the rule-based approach uses a template-matching of the nose bridge. In addition, the stochastic algorithm applies features derived from the optical flow, and classifies them by a set of discrete Hidden Markov Models. The rule-based implementation evaluates the key-feature in a finite state machine. We extensively tested the systems in two different application domains (VR desktop scenario vs. automotive environment). Six different gestures can be classified with an overall recognition rate of 93.7% (rule-based) and 97.3% (stochastic) in the VR (92.6% and 95.5% in the automotive environment, respectively). Both approaches work independently from the image background. Concerning the stochastic concept, further gesture types can easily be implemented.