ROBUST HUMAN MOTION RECOGNITION FROM DISTORTED WIDE-ANGLE IMAGES FOR VIDEO SURVEILLANCE
Daisuke Miki, Shinya Abe, Shi Chen, Kazuyuki Demachi · The Proceedings of the International Conference on Nuclear Engineering (ICONE) · 2019
Installation of surveillance cameras in nuclear power plants (NPP) is critical to protect the facilities against terrorist attacks, bombings, or sabotage. This has led to the generation of huge amounts of surveillance video data, giving rise to a demand for a technique that can automatically detect anomalies or suspicious movements. Tracking human motion from video sequences is a notable technique used for detecting anromalies in human behaviour. Motion-capture devices recognize human motion using a depth camera. However, the use of a depth camera incurs a problem in that a complicated camera system is required, and the angle of view is limited. To overcome this problem, there is a need for a means of recognizing human motion in wide-angle images. In this study, we devised a method for tracking human motion that is robust to wide-angle image distortion. The main contribution of this study is a methodology that automatically estimates the transformation parameters needed to improve the accuracy of motion recognition; these parameters are applied to a distorted wide-angle image in every frame. We propose a new multi-layered convolutional neural architecture for estimating the locations of human joints in images and transformation parameters simultaneously. The robustness of our method when applied to distorted wide-angle images is demonstrated through a quantitative evaluation of human joint prediction. In addition, we compare our method with a motion tracking system and an infrared-camera-based motion capture system to demonstrate its ability to handle wide-angle and close-range images.