Smart Home Surveillance System

Abstract

The Smart Home Surveillance System is an advanced intelligent security system that makes our residence safe and secure. The system aims to transform the regular CCTVs into intelligent cameras that will have some additional features like Suspicious Activity Detection, Smart Video Recording, Object Detection, and Notification alerts. The system is built not only for residential security but also for commercial areas. In fact, it can be installed in any environment either working or residential to track and minimize the occurrence of criminal activities. It detects the activities like Robbery, Assault, Stealing, Burglary, and Wall Climbing. The system's mobile application uses a trained deep learning model (CNN) that classifies the activities as suspicious or non-suspicious. The model has trained on the UCF crime dataset as well as our custom-made dataset, based on which the binary classification is being done. If the activity is classified as 'Suspicious', the user will get notified immediately on the app through the snapshot of the scene. Yolov5 has been used for object detection. Further, there's a smart recording option that can be enabled through the mobile application. By enabling the smart recording option, the system will start recording only on the detection of objects in the frame. There's an admin portal of the system, which is a web app that would be used by the administrators. The user will be assigned the login credentials through this portal, and it is also responsible for monitoring the whole system. There is also a web portal on the user side that shows the live streaming once the user logged into the portal through the provided credentials. Thus, the whole system facilitates the users by providing a secure and intelligent security system. To evaluate the model, the confusion matrix has been used. The confusion matrix is being created using a test set which is obtained from the UCF crime dataset and some custom recorded videos. The model has an accuracy of 89% with the best confusion matrix obtained so far.