An Ant Learning Algorithm for Accelerometer-based Gesture Recognition

Abstract

Today’s emerging gesture recognition techniques have enriched the ways of human machine interaction. With the popularity of smart devices such as iPhone and iPod Touch, accelerometer-based gesture recognition for facilitating such interactions is becoming even more pervasive and promising. Accelerometer-based gesture recognition systems have been extensively discussed in many previous related work. Currently, there are several techniques being applied for recognizing gestures, most well-known algorithms are Hidden Markov Model (HMM) and Dynamic Time Warping (DTW). However, they do have shortcomings: 1) HMM requires a sizeable amount of training data, and suffers from the high computational overhead for both training and classification. 2) The processing time of DTW depends on both the length and number of templates.

In this thesis, we introduce a novel gesture recognition algorithm named the Ant Learning Algorithm (ALA), which aims at addressing some of the limitations with the currently two leading algorithms, especially HMM. It takes advantage of the pheromone mechanism from ant colony optimization and uses pheromone tables to represent gestures, which scales well with gesture complexity.

ALA requires minimal training instances and greatly reduces the computational overhead required by both training and classification. The experimental results show that ALA can achieve a high recognition accuracy of over 90% with only one training instance and exhibits good generalization.