Binary classification of floor vibrations for human activity detection based on dynamic mode decomposition

Analyzing small amplitude of floor vibrations is a new promising means for identifying the types of human activities, e.g., walking around and accidental falls. In this paper, we consider the binary classification problem of floor vibrations for the applications like fall detection. For practical use, there are two main issues of the problem. First, the prediction of the classifier should be fast. Second, the training set is sometimes small and the diversity of negative samples brings extra challenges when the training samples are insufficient. The state-of-the-art methods for time series classification, such as HIVE-COTE and ResNet, are computationally intensive or susceptible to the size of the training set. Therefore, we propose a new feature extraction method based on dynamic mode decomposition (DMD) and high-frequency characteristics, whose time complexity is linear with the size of training set and quadratic with the length of time series. The method is evaluated on the dataset of floor vibrations proposed by Madarshahian et al. (2016). The results show higher accuracy compared to the ResNet classifier and time series forests, especially when the negative training samples are deficient in type.