A Survey on Unusual Event Detection in Videos

Abstract

As the usage of CCTV cameras in outdoor and indoor locations has increased significantly, one needs to design a system to detect the unusual events, at the time of its occurrence. Computer vision is used for Human Action recognition, which has been widely implemented in the systems, but unusual event detection is lately entering into the limelight. In order to detect the unusual events, supervised techniques, semi-supervised techniques and unsupervised techniques have been adopted. Social force model (SFM) and Force field are used to model the interaction among crowds. Only normal events training samples is not sufficient for detection of unusual events. Double sparse representation has been used as a solution to this, which includes normal and abnormal training data. To develop an intelligent video surveillance system, behavioural representation and behavioural modelling techniques are used. Various machine learning techniques to identify unusual events include: Graph modelling and matching, object trajectory based, object silhouettes based and pixel based approaches. Kullback–Leibler (KL) divergence, Quaternion Discrete Cosine Transformation (QDCT) analysis, hidden Markov model (HMM) and histogram of oriented contextual gradient (HOCG) descriptor are some of the models used are used for detecting unusual events. This paper briefly discusses the above mentioned strategies and pay attention on their pros and cons.