Real-Time Activity Recognition for Surveillance Applications on Edge Devices

Abstract

Human Activity Recognition is a crucial task for surveillance systems that has seen great advancements with the emergence of Artificial Intelligence. At the same time, hardware advances have allowed for development of systems that operate in real-time. However real-time performance is still a far from solved problem for wearable devices when it comes to computer vision tasks such as activity recognition. In this paper a hybrid solution for Human Activity Recognition is proposed that exploits a lightweight method for on-device posture recognition and a more heavyweight activity recognition method executed on the cloud. The experimental evaluation for the activity recognition module indicates superior performance compared to existing methods and the lightweight posture method can predict satisfactorily the desired classes. The developed system offers a user-friendly Augmented Reality application that provides scene annotations to the user including the activity of the detected persons.