부산대학교 도서관

Nowadays the research that abnormal events is automatically recognized in video surveillance system has been proceed. However there are some issues that are difficult to solve, typically occlusions issue in video stream including crowd events and another issue of restrictive training dataset of normal/abnormal events. To solve these problem, we propose a new framework for abnormal events recognition in crowd scenes. The proposed method consists of the features extraction module and the recognition module using that features. The moving energy and the stationary energy in video streams are extracted by the features extraction module. And normal/abnormal events are classified by the recognition module that use modeling based on HMM and an algorithm to calculate an optimal threshold value. Our proposed method is showed by an experiment using an synthetic dataset and an actual video dataset.