A Lightweight Asynchronous Fusion Framework for Multimodal Surveillance Anomaly Detection

Multimodal anomaly detection has gained significant attention in modern surveillance systems due to its ability to integrate information from multiple data sources and improve situational awareness. Existing approaches commonly combine modalities such as video, audio, thermal imagery, and sensor data to enhance anomaly detection performance. However, many current fusion techniques rely on strict temporal synchronization and assume that heterogeneous data streams are temporally aligned. In practical surveillance environments, modalities often operate at different sampling frequencies and generate data asynchronously, creating temporal inconsistencies not adequately addressed by many conventional fusion approaches. Furthermore, existing asynchronous learning methods frequently employ computationally intensive architectures that may limit their applicability in resource-constrained surveillance settings. This paper reviews recent developments in surveillance anomaly detection, multimodal fusion, temporal alignment, and asynchronous learning. Based on the limitations identified in existing literature, a lightweight asynchronous fusion framework is proposed for multimodal surveillance anomaly detection. The framework utilizes independent feature extraction, temporal windowing, and asynchronous fusion to facilitate the integration of heterogeneous surveillance data streams without requiring strict temporal synchronization. By emphasizing simplicity, flexibility, and practical deployment considerations, the proposed framework provides a conceptual foundation for future research on lightweight asynchronous multimodal surveillance systems.