A feature evaluation method for template matching in daily activity recognition

In daily activity recognition system based on inertial sensors like accelerometer, different features can represent raw data in different aspects. In order to extract the most proper ones among so many features, we propose a feature evaluation method for template matching. Our method uses the dynamic time warping method to calculate the distances between features in training phase, thus could choose the most proper feature among various ones and highly reduce the complex of calculation since all of the work will be done in training phase. We use this method to evaluate 6 commonly used features on the OPPORTUNITY Activity Recognition Datasets which includes 17 classes of activities. The experimental results show that our proposed method has high positive relationship with the real accuracy tested by testing data, thus demonstrate that our proposed method is efficient.