Abstract:
Security today is a sensitive issue. At present in MCS, surveillance is done manually but with our application developed and installed a lot of stuff can be automated. Our algorithm will be generalizable so given enough data and equipment it could be generalized for detecting other suspicious stuff in the video. It is a key step towards an era of surveillance.
Video Analysis System is a python based desktop application which is used to analyze live camera feed to detect the anomalies before hands. It includes cross platform support and can be executed on any three of known Operating Systems.
VAS will take video feed as input and will tell (through interface) what objects are existing in current frame. A suspicious object includes a person trying to breach a wall or a tool i.e. knife. This has been done using a deep learning algorithms which is being trained on 4 multiple datasets. First dataset was being made using a CCTV camera on the wall of MCS, Second Dataset was being scrapped from Internet (due to insufficient info in Custom dataset) Images of weapons were being scrapped from the internet. The third dataset was made using a mobile phone with intention to classify any person who is trying to breach the wall. Last dataset used was MS-COCO, which is a labeled dataset for almost 80 different objects Using these datasets, we almost made 8-10 deep learning models. Including the renown architectures such as Mobile-Net, YOLO, Simple CNN. The final product consists of two working models which classifies all classes other than weapons.
