SAFE—A Scalable and Agile Framework for Mixed Critical Event Detection

Real-time video stream processing is challenging for detection of mixed critical events. Challenges lie in developing a scalable platform that not only detects multiple events according to their criticality but can also meet multi-tenancy requirements. Significant decisions such as monitoring rate and modality of the computer vision model can not only impact the robustness of the system but can also downgrade the scalability of the system. In this paper, we cater this challenging problem. We propose SAFE — a Scalable and Agile Framework for Event detection. SAFE utilizes a stateless deep learning model to support multiple users and incorporates adaptive frame rate monitoring, which prioritizes high-severity events while reducing the processing of low-severity events, to lower cloud costs. Through extensive experiments on different critical events (such as fire and traffic), we evaluated SAFE with respect to accuracy, scalability, low latency, and robustness. We demonstrate the efficacy of our system and validate its usage for a real-time mixed critical event detection system.