Abnormal crowd event detection based on outlier in time series

Abstract

Crowd management research shows a lack of depth in the literature insofar as most major incidents can be prevented or minimized by a proper management strategy. Specifically, if abnormal crowd events can be detected early and the relevant governing agency can take appropriate actions towards mitigating the dangers, accidental injury can be prevented or the incident can be contained. This paper presents a technical approach to gather the required crowd data using fixed cameras to collect visual data while using a grid model to describe the crowd distribution. The measured area will be divided into several unit areas and each unit area is considered to be a simple cell in a grid model. The state value of each unit area is determined by changes in the total number of active pixels within the unit area. Under the circumstances, the motion status of the measured area is represented by a dynamic state matrix, which will save computing time. Should abnormal crowd events develop, a crowd tends to attempt to quickly leave the area and the resultant crowd distribution changes accordingly. Thus, we expect the crowd distribution to have a sudden and significant change as the abnormal crowd event unfolds. According to the Uniqueness Theorem, crowd distribution in the grid model can be described by a series of different order moments. To consider the normalization, the Chebyshev moment is used to describe the status distribution in the grid model of each image and a time series is used to describe the varying features of the two adjacent images in the video data. When unusual events occur, the status distribution in the grid model between images will show more obvious changes within a short period of time. We can use an outlier detection method to detect the extreme values and also identify the type of outlier: additive outliers or innovational outliers are used to determine the cause of abnormal feature variation.