Hand Gestures Recognition Based on Lightweight Evolving Fuzzy Clustering Method

Robots of the future will socialize with humans. Human-robot interaction (HRI) by a vision-based gesture interface helps to personalize the communication with humans in various contexts - from support of their daily life to social skills training of children with developmental problems. We are especially interested in vision-based hand gesture HRI and propose a hand gesture recognition system based on a novel online extraction and classification scheme, which is lightweight and can be used in a mobile robot. An online Lightweight Evolving Fuzzy Clustering Method is used to categorize the positional and HSV model of pixels for the edges of the gesture image. The result clusters consist of (x, y) coordinates and the averaged grayscale level at these locations. Then these clusters are processed to identify typical for the hand features brighter and darker pixel information. The database consists of averaged grayscale levels in HSV format for neighbor pixels that characterize different features. For feature recognition we use Tanimoto similarity measure for matching the current grayscale patterns to those in the database. Then the feature location is encoded in a binary format. For gesture recognition we use a formalism of Symbol Relation Grammars to describe a gesture, as well as simple and fast bitwise operations to find the position and orientation of the features in the gesture.