Multimodal Anomaly Detection for Urban Safety: A Real-World Implementation in Large-Scale Surveillance Systems

Abstract Effectively identifying abnormal events in urban environments requires the fusion of diverse data modalities, including video, audio, and environmental sensors. This paper presents a neural network-based framework for real-time multimodal anomaly detection deployed in a live metropolitan scenario. Using a combined CNN-LSTM structure and attention-based feature alignment, the system integrates heterogeneous inputs to classify anomalies across 57 event types. Field deployment across 28 surveillance zones demonstrated a detection precision of 93.4% and average processing latency of 0.89 seconds per frame group. The results validate the feasibility and robustness of deploying neural models for large-scale, cross-modal urban safety tasks.