Integrating gesture-based identification in context-aware applications: a system approach

Gestures contain enough information to distinguish a person from another. This `biometric' feature is the basis of the Kinect-enabled system described in this paper, which is designed to identify a given person through a short chain of in-air gestures (2-3 ones). The user-dependent (trainable) system relies on a traditional Dynamic Time Warping classification algorithm. This core algorithm enhances its performance thanks to two filters (duration and sensitivity) that make the system more robust against the variability of the gestures from the same user among iterations (e.g. gestures may differ in duration and shape). The system includes a `point-based' score strategy designed to rapidly discard the non-probable users and to speed up identity confirmation among the most probable candidates. Additionally, the system includes a specific component to guarantee the training quality. In order to integrate the identification process into a real service, the gesture chain is retrieved through a sequence of double-option questions that aim at better adapting a smart space (the lighting infrastructure and the available multimedia contents) to a given user depending on his specific needs. This way, the identification task remains in second plane, wrapped in functional questions. Ten users have tested the service, providing their positive feedback.