Online data segmentation based on clustering algorithm and autoregressive model for human actions recognition

Recognition of human actions by using wearable sensors has become an important research field. Segmentation to sensor data is a vital issue in reconstructing and understanding human daily actions, and strongly affects the accuracy of human actions recognition. Traditional online segmentation approaches are mostly designed for one-dimensional sensor data, which greatly limits these approaches to multi-dimensional wearable sensor data. In this study, an online data segmentation approach based on clustering algorithm and autoregressive model (AR) is proposed, which can dynamically choose suitable dimensions. First, rough classification is done by clustering algorithm. Then, ARs are used to determine the changing point of different human actions. Precision, recall and F-measure are introduced to evaluate the segmentation results. The experimental results demonstrated that the proposed method outperforms some existing approaches, including HMMs, adaptive models and fixed-threshold method. By using the proposed method, the accuracy of human actions recognition reached 86.5% against ground-truth, which was better than other methods mentioned in this paper.