広角映像の歪みに頑健な注目点検出手法の開発と人物動作解析への応用
概要
Tracking human motion from video sequences is a notable technique that is used to detect anomalies in individual human behavior. Several commercially available motion capture devices are based on the use of depth cameras. However, there are a couple of problems with the use of a depth camera. Firstly, a complicated camera system is required, and secondly, the optical field of view is limited. To overcome these problems, we need a technique that can recognize human motion from wide-angle images. In this study, we will devise a method for tracking human motion that is robust toward the distortion of wide-angle images. The main contribution of this study is the development of a methodology that can automatically estimate the transformation parameters that are required to improve the accuracy of human motion recognition. We propose a new architecture of a multi-layered convolutional neural network that can estimate the location of human joints in images and transformation parameters simultaneously. We confirmed its applicability to human motion analysis by comparing the results of the application for both natural and unnatural human motion data.