Adaptive gesture tracking and recognition using acceleration sensors for a mobile device

Abstract

We present in this paper an adaptive gesture classifier for mobile devices, along with an efficient method to automatically detect endpoints of gestures. A classification model based on 1- NN with DTW-based k-means clustering is augmented by a metacognitive framework that measures the quality of the learned model and continuously updates it to improve the performance. We evaluated the model with an accelerometer signal database of 26 English alphabets. The results showed that the adaptive framework improved the recall and precision rates by 4.9% and 5.6%, respectively. Our endpoint detection method, based on energy variance and low-pass filtering, successfully detected 98.5% of gestures with an average detection delay of 176 ms.