Abnormal Event Detection in a Surveillance Scene Using Convolutional Neural Network

Introduction

Nowadays, the demand for security of people and personal properties is constantly increasing, so video surveillance has become major daily anxiety. The prominence of this demand has led to the placement of cameras widely, which produce a large quantity of video. Most existing video surveillance systems are fully supervised by humans. Video monitoring is a very cumbersome and time-consuming task. A human can't find out the abnormal events from large size video. However, even one small mistake could cause an unacceptable loss. Thus, it is essential to develop a system dealing with many video frames and alert people for a punctual and functional response when an abnormal event occurs. So, Lots of research on automatic video surveillance is going on. In this research, a crowded environment is considered for surveillance scene, and the focus is to detect an abnormal event in a crowd scene. Abnormal event detection is regarded as one module of crowd analysis (Julio Cezar Silveira Jacques, Jr. et al, 2010). Recent research (Cem Direkoglu et al, 2017) reports that 30% of the research work of complete crowd analysis is done in the abnormality detection module.

Automatic abnormal event detection in a crowd scene is challenging because the definition of abnormal is subjective or context-dependent. The event which occurs rarely can be considered as an abnormal event. Abnormality detection is the problem of intricate sequential visual patterns' recognition, as the crowd scene contains occlusion and clutter. The conventional techniques use standard low-level features like optical flow and Histogram of Oriented Gradients for this task, which are ineffective at identifying complex patterns and hard to implement in real-time applications. All types of approaches, like supervised, unsupervised and semi-supervised learning, are used by researchers to detect an abnormal event in a crowd scene.

Convolutional Neural Network gives excellent performance on various computer vision areas like video summarization, behaviour recognition, security, object tracking etc. Inspired by the performance of CNNs in the mentioned domains, the authors have proposed a CNN based approach to detect an abnormal event in a surveillance scene. The proposed method is based on supervised learning. It is impossible to generate labels of all types of abnormal behaviours for classification, so general-purpose abnormality detection may not be possible. However, as per requirements at a particular place, definitely supervised learning can be used. For example, in some areas, only pedestrians are allowed, so vehicle entry is abnormal. At some paid entry points, access without making a payment is considered deviant behaviour. At some places like bank or shops, robbery can be considered as an abnormal event. For surveillance cameras positioned on the road, the events like suddenly running, fighting, accident, crowd formation can be considered abnormal. In an examination hall, talking or copying the answer from the other student's answer sheet is abnormal behaviour. So, as per the situation, it is possible to generate labels of abnormal events and supervised learning can be used. Many real-world datasets for abnormal event detection in a crowd scene are currently publicly available, as it is an open research problem. In this work, four benchmark datasets are used, namely UMN (University of Minnesota), UCSDPed1(UCSD Anomaly Detection Dataset, 2013), Violent Flows (T. Hassner, et al, 2012) and Subway Entry (A. Adam et al, 2008), in which different anomalous behaviours are recognized like running, use of vehicles, walking across walkway, violence, moving in a wrong direction and avoiding payment. The CNN based techniques which already exist either use a pre-trained CNN model or apply lots of efforts for pre-processing. The critical contributions of the proposed method are summarized as follows: