Anomaly Detection in Video Surveillance

Abstract

In this project, we have developed an anomaly detection system which makes use of Machine Learning to detect anomalies to include Violence, Theft, Accident, Arson, and Abuse. This would be accomplished by using deep neural networks. The approach adopted to fulfill the requirement is Multiple Instance Learning approach that considers normal and anomalous videos as bags and video segments to be the instances. Thus automatically learning an anomaly model to predict high score for anomalous video segments. The training datasets consist of a variety of videos containing normal and anomalous (Explosion, Shooting, Road accident and ten other anomalies) of approximately 128 hours containing 1800 real world surveillance videos. After the training phase, Model is then deployed using interface which takes the video as an input and displays results as graph. The Summary of anomaly detected further displayed in a GUI containing anomalous frame, threshold, mean and standard deviation. In addition to this the system has access control mechanism in the form of login and maintaining logs. The system is also used for trend analysis that will help security personnel to enhance security on ground. Hence the system provides management solution for video surveillance.