Real-time Hand Gesture Identification for Human Computer Interaction based on ICA of Surface Electromyogram


Dinesh Kant Kumar.

Student: Ganesh Naik.

Hans Weghorn, University of BA, Stuttgart, Germany.

Introduction: Today, there exists a variety of interfaces that allow human users to interact with computerized systems. Many of these input and output devices force the user to adapt to the requirements of the machine construction, like e.g. numeric keyboards on tiny devices often have to be used also for letter input.
Significance: In contradiction to such technically-driven concepts, the aim of the investigation presented here is to provide a reliable input mode, which enables machine control for rehabilitation and human-computer interaction applications in a quite natural way. The processing in this new input system consists of three major stages: At first, hand gestures are sensed from non-invasive surface electromyograms, and in the second step the activities of the involved individual muscles are decomposed by semi-blind independent component analysis (ICA). In the last step, the particular hand action is identified with an artificial neural network (ANN).
Applications: These experimental results demonstrate that the proposed approach yields a high recognition rate with various gestures, and the system was verified being insensitive against electrode positions. A comparative evaluation of applying the same recognition mechanism in identifying facial movement yields new findings about the properties of the derived ICA mixing matrix, which can be exploited as indicator for the reliability and efficiency of the pattern classification mechanism in a distinct application.
Challenges: In this model-based approach, the order and magnitude ambiguity of ICA have been overcome by using a priori knowledge of the hand muscle anatomy and a fixed un-mixing matrix for the signal decomposition. In 360 single-shot experiments, this system was able to classify all tested hand gestures fully correct.
top of page