Accurate 3D human activity recognition and recursive bayesian fire recognition for intelligent video surveillance

Abstract

Automatically detecting abnormal behaviors and situations are known as important problems for in-telligent video surveillance. Human activity recognition and detecting fire using video analysis are key tech-nologies for the detection of the abnormality. In this paper, I attempt to achieve high accuracy in the detec-tion of the two major problems. For the human activity recognition, existing methods have tended to re-search center around 2D-based and view-dependent representations, which make the accuracy of the meth-ods lower. I adopt a 3D activity modeling and a representation method view-invariant and rotation-invariant and try to improve the recognition accuracy. For the detection of fire, existing methods have tended to utilize strong assumptions about fire characteristics, which make the method not applicable in real environments. To utilize and incorporate natural characteristics about fire in space and time, I adopt a recursive Bayesian estimation and greedy margin-maximizing clustering. With the methods, I achieved high accuracy recognition and detection results compared with existing methods, which are proved in various experiments in various situations.