Automated Surveillance Model for Video-Based Anomalous Activity Detection Using Deep Learning Architecture

Abstract

Due to need of monitoring and safeguarding the activities of people and achieving global security, surveillance cameras are deployed in public places. It is important to find out interesting patterns such as suspicious behavior of entities (people, objects) from such videos. Automatically monitoring such videos for detection of anomalies without any assistance of human personnel is the need of the hour. Therefore, deep learning approach for modeling video-based anomalous activity detection has been put forth in this paper. The problem of anomaly detection has been handled as one-class classification problem. The proposed methodology involves auto-encoder-based two-dimensional convolutional neural network for feature learning, long short-term memory model for identifying temporal statistical correlation and radial basis function as activation function in fully connected network for one-class classification. Use of deep learning model for calculating the likelihood from multivariate Gaussian distribution has made our approach a novel one. The proposed methodology perfectly works for simulated data being generated for anomaly detection, and this justifies that it is the best suitable methodology for modeling video-based anomaly detection.