Video-Analysis-as-a-Service with Foundation Models for Real-Time Multi-event Detection

The increasing deployment of camera networks in smart city environments has increased the need of video surveillance systems able to detect heterogeneous events in real-time while operating under resource-constrained conditions. Traditional video analytics approaches usually rely on task-specific models, which require dedicated training data and architectural redesign. In this work, we propose a unified framework for real-time multi-event detection based on foundation models for video-text matching. The proposed approach represents short video segments and textual event descriptions within a shared embedding space, enabling event detection through cosine similarity between video and text embeddings. The framework is implemented within a Video-Analysis-as-a-Service architecture, where the Open Vocabulary Event Detection (OVED) service acts as a modular semantic engine invoked on demand for different surveillance tasks. The proposed method is evaluated across three surveillance scenarios, including fall detection, fight recognition, and illegal waste dumping detection. Experimental results show that the framework achieves competitive or superior performance compared with specialized task-specific methods, while requiring only the modification of textual queries to address new events. Moreover, the proposed approach remains feasible on constrained hardware, confirming its suitability for scalable edge and cloud deployments.