Using Cultural Algorithms to Improve Wearable Device Gesture Recognition Performance

Abstract

Wearable computing devices are now mainstream. Many such devices have capable MEMS sensors that can be exploited for recognizing dynamic, in-the-air gestures. Noting the somewhat limited compute and battery life of today's devices, we present a computationally efficient approach to gesture recognition that can be effectively used inside an app running on standard, off-the-shelf hardware, such as an Android Smart watch. Our approach has two phases. The first phase is to construct finite state machines (FSM) for gesture recognition. We present a novel approach - which leverages techniques from functional programming languages - to define rich yet compact FSM. Such FSM can be further tuned for higher accuracy with help of some training data and a suitable optimization method - this is the second phase. To demonstrate effectiveness, we created a gesture recognition system for an automotive scenario as an Android Smart watch app and then tuned the gesture recognition engine using Cultural Algorithms optimization and training data. We achieved 77% gesture recognition accuracy which is on par which more computationally intensive techniques such as Hidden Markov Models.