Human Posture Classification from Multiple Viewpoints and Application for Fall Detection

Abstract

Vision based human detection and posture classification are essential components of many computer vision systems in security, advertisement and healthcare. Although human detection has been studied for more than three decades, most of proposed methods focused on standing people. In a daily-life applications, posture of human can vary strongly from standing, sitting, lying to crunching. In addition, human shape also vary according to different viewpoints. This makes more challenges for human detection and posture classification. Due to a smooth transition of postures, it is difficult to determine how many postures should be considered and classified. In this paper, we deploy an unsupervised technique to explore the number of distinctive human postures from a given set of activities. We then resolve the problem of postures classification as a multi-class detection problem using a state of the art convolutional neural network YOLO. The proposed method gives promising results on a dataset of human activities taken from six views. Recall and precision of human posture detection and classification are highly achieved on every viewpoint (99% of recall and precision). The posture classification leads to potential application for fall detection at high frame-rate.