Anomaly detection in videos

Abstract

With the development of modern digital cameras and CCTVs, our cities and their important public places can now be monitored 24x7 around the clock. Traditional video analytics methods necessitate ongoing surveillance monitoring, which is time consuming. A lot of human resource is needed for running such systems to observed and react to any abnormal event. Hence there is a need to develop automatic video anomaly detection systems to reduce the dependency on human resources. An anomaly can be described as an extraordinary event or an emergency that differs from the norm. Finding and classifying anomalies in videos is known as video anomaly detection. This thesis discusses the various algorithms and techniques that are used to create anomaly detection systems. We have proposed a model using Multiple Instance deep learning network for solving the problem of video anomaly detection. We have used the two stream Inflated 3D Convolutional Neural Network (I3D) for feature extraction from the RGB and optical flow data stream. We propose a modified loss function based on the deep ranking loss criteria to improve the model�s effectiveness. For training and testing the model, we have used the UCF Crime dataset. To check the model�s effectiveness, we have used the Area Under Curve (AUC) value of the Receiver Operating Characteristic (ROC) curve and compared it with the state of the art methods We have also compared the loss resolution of the standard ranking loss function with that of our modified loss function. Finally we have compared the anomaly activity classification accuracy of our proposed model with that of the state of art models.