Abstract: In the study of human activity recognition(HAR) based on the inertial sensor, feature extraction was one of the key points. The stability of discrete data statistical features depended on the window size of feature extraction. Generally speaking, the length of window needed to be greater than one motion cycle. Therefore, compared to the traditional behavior recognition, short-time behavior recognition was more difficult. Thus a novel template matching method was proposed for recognizing the test samples whose durations were shorter than one motion cycle. Firstly, by properly dividing the long sequence samples, a complete short-time activity template library was constructed. The short-time samples to be tested and the samples in the template library were processed and matched. Secondly, in the matching algorithm, the similarity histogram was obtained by utilizing the sum of the F norm distance between the samples and the 2 norm distance of the global gradient vector as the matching metric. Finally, based on the similarity histogram, the final classification recognition results were obtained according to the voting strategy. Experiments show that in the case of using a single sensor to identify short-term behavior, the new algorithm had better performance than traditional algorithms in accuracy and stability, and can be adapted to short-term behavior classification problems under different windows.