Low-Level Human Action Change Detection Using the Motion History Image

Abstract

Human action recognition is an active topic in computer vision. In recent years, tangible results have been shown through deep learning methods; however, at a very high computational cost. They may be suitable for video retrieval or video summarization of a movie or drama, but they are not suitable for the visual surveillance of human action, which should be analyzed in real time. In this study, we propose an action change detection method to reduce the computational cost. On the surveillance camera screen, anomalous actions such as movies and sports are frequently not observed, and simple actions are often repeated. Thus, it is very inefficient to continue to apply high cost action recognition methods on repeated actions. The proposed action change detection method decides whether the previous action of the person is the same as the current action. The action recognition method is applied only when it has determined that the action has changed. The action change detection process is as follows. First, extract the motion history image from the input video and create one-dimensional time-series data. Second, determine the action change using the change of time-series data trend by the threshold. Experiments on the proposed method were performed on the KTH dataset.