Real Time Abnormal Activity Detection using Machine Learning Algorithms for CCTV Surveillance System

Abstract

The demand for security and reassurance to the public has always been the foremost concern of the society. Anomaly detection plays a prime role in surveillance applications. It is one of many enabling technologies for increasing security, enabling law enforcement and other security personnels to instantly respond to potential threats. In this thesis, the issue of identifying any abnormal event in the surveillance domain has been studied, with a literature review that identifies some weaknesses in previous anomaly detection techniques. Constant observation of surveillance cameras generating sheer volume of data by humans is merely unfeasible. This inconvenience is creating a dire need to accurately automate the entire process. When threats are definable, we can use methods based on situation recognition to detect them, but sometimes the anomalies are hard to define. In such cases a technique called data-driven anomaly detection is applied. In data driven anomaly detection a model of normalcy is trained and utilized to find anomalies. Anomalous activities are then alarmed providing instant intervention and prevention of criminal activity dispensing an effective and efficient surveillance and reducing the labor of constant monitoring. We intend to utilize Deep Learning (DL) model: “Autoencoders” to identify and classify activities from a batch of 10 video frames. The evaluation of proposed solution is done on data sets, particularly for corridor and indoor settings. Conclusion of the thesis is that the proposed model is strong and suitable for use. Keywords: Security; CCTV cameras; Deep Learning; Spatio-Temporal Data; Prediction; Machine learning, Convolutional neural networks, Image processing.