A novel method for short-time human activity recognition based on improved template matching technique

Traditional human activity recognition algorithms based on wearable sensors have a drawback in recognizing short-time samples. Due to, the instability of the traditional time and frequency features in short-time sample condition, the recognition results can be seriously affected. This paper proposes a new algorithm for human activity recognition based on improved template matching (ITM) for the short-time samples. Our proposed approach is based on four stages: First stage, we transform all longtime samples into short-time activity template with sliding window and structure over-completes training template set. Consequently, each kind of pattern of the activity will contain adequate atomic patterns. Second stage, we apply the short-time activity template instead of the unstable traditional features to describe the actions, where each activity template represents one short-time kinestate of the activity. Third stage, matching the short-time test sample with the over-complete training template set directly and calculate the residual between the test sample and each training template. Last stage, we determine the label of the test sample according to the smallest residual. In this paper all of the public WARD1.0 and our database are used to show the robustness of the proposed method under the conditions of using short-time samples (i.e., about 0.3s). In particular, the recognition rates which are based on the above two databases can reach to 96.1% and 96.8% respectively.