Video summarization for anomaly detection using deep learning

Abstract

Video summarization creates a concise summary of the content of a lengthier video document, by selecting and presenting the most helpful or intriguing components. It has wide range of applications in various domains, like surveillance, entertainment, education, advertisement and more. In the current work, we propose video summarization of anomalous content in surveillance videos. Crime has become a concerning problem in today’s evolving world. Close scrutiny of our daily visiting places, be it house, office, malls, etc., has become a necessity for us with crimes increasing every day. This inspection is being achieved by installing surveillance cameras everywhere. The videos recorded this way comes in handy for crime investigations. But the problem is to go through the complete recorded video to find if it has some anomalous content, which is both time and effort-consuming. This brings in a requirement for some mechanism that can save our resources without compromising the information. A similar approach is to summarize the anomalous content of a video in terms of a single image. To be more precise, we need a 2-step pipeline, i.e., a system which can precisely identify the abnormal content in a video and then combine this extracted content into a single image. In this project, we try to build a deep learning model which can identify anomalous frames and a rank pooling algorithm that can combine these anomalous frames into a single dynamic image.