Real Time Gathering Event Detection Based On Layered Hidden Markov Model

Video-based event detection is one of the most promising applications of computer vision. Its purpose is to let the machine recognize the interested video event automatically. Gathering event is a kind of abnormal event achieved by multiple agents in a complex video scene. The detection of gathering event is still a challenging task.In this paper we study the problem of gathering event detection. Especially focus on the description and the modeling problem of the gathering event.After analyzing the existing methods for video event description, this paper introduces a scene-division-based method to describe a gathering event. We propose a method to divide the video scene into blocks of area. Then the number of people and their moving speed are extracted from each block as the features of the video event. This method gives a good description of the gathering event and also avoids tracking individuals in a crowded scene. In the mean time, we analyze the popular optical flow method for speed evaluation. Since the optical flow method is too complex in computation and is sensitive to environmental interference, a'running average image'method is used to evaluate speed instead. The introduction of this method significantly reduces the computation and brings robustness to the feature extraction procedure.Since the popular Hidden Markov Model (HMM) is too sensitive to model a gathering event and is easy to overflow, this paper uses Layered Hidden Markov Model (LHMM) to model gathering event. The introduction of LHMM greatly improves the system's capability to analyze complex video event. It also improves the system's reliability and reduces the false detection rate.After an overall evaluation, the experimental results show that: by using the proposed method to describe and model gathering event, the system can detect the gathering event in real-time with high detection rate, low false detection rate, and with robustness.