Accelerometer Gesture Recognition

Our goal is to make gesture-based input for smartphones and smartwatches accurate and feasible to use. With a custom Android application to record accelerometer data for 5 gestures, we developed a highly accurate SVM classifier using only 1 training example per class. Our novel Dynamic-Threshold Truncation algorithm during preprocessing improved accuracy on 1 training example per class by 14% and the addition of axis-wise Discrete Fourier Transform coefficient features improved accuracy on 1 training example per class by 5%. With 5 gesture classes, 1 training example for each class, and 30 test examples for each class, our classifier achieves 96% accuracy. With 5 training examples per class, the classifier achieves 98% accuracy, which is greater than the 10-example accuracy of other efforts using HMM’s[1, 2]. This makes it feasible for a real-time implementation of accelerometer-based gesture recognition to identify user-defined gestures with high accuracy while requiring little training effort from the user.