Accelerometer Based Real-Time Gesture Recognition

Abstract

Gesture is a natural expression form for humans, but its recognition is a similarly hard problem as speech recognition. In this paper, I present a real-time hand gesture recognition system, which identifies relevant parts in the continous sensor data stream, and classifies them to the most probable gesture. Instead of the usual button-based segmentation, I have created an automatic segmentation method, which makes the interface more natural. The results showed that the two different classifiers reach 97.4% and 96% accuracy on a personalized gesture set, and these results can be improved for certain gesture sets with the combination of the two algorithms. Furthermore, the system has great performance and low response time, so the user experience is much better than with previous gesture recognizers.