Crowd Abnormal Behaviour Detection using Convolutional Neural Network And Bidirectional LSTM

Abstract

In the present day, for the purpose of social security, video surveillance frameworks are utilized in public places such as metro stations, temples, shopping malls, stadiums and so on. However, these video monitoring systems are merely utilized as a recording framework that cannot immediately identify and detect an abnormal event, and it is difficult for an individual to ensure a steady eye on the screen. The goal of the proposed work is to create a model that can classify normal and abnormal crowd Behaviour using a real time video surveillance system to detect abnormalities and monitor congested metropolitan areas. Because the recorded film is strong proof against the offender, crowd analysis will assist security services in stopping crime, thus helping in keeping crime under control. The existing works contain high computational cost and the present datasets are not sufficient to learn the complexity of real- time video surveillance data. These factors motivate us to create a model that requires less training time as well as a custom dataset with sufficient training samples. This method uses an efficient deep learning model CNN+BiLSTM for detecting and classifying abnormal Behaviour from normal. The proposed model gave an accuracy of 98.5 for our custom dataset. The custom dataset contains video clips, which covers different scenarios of crowd havoc and panic for training and testing our model.