Identification of urban fire safety risks based on multimodal data fusion algorithm

Abstract

Aiming at the weak anti-interference performance of monitoring systems and equipment, frequent false positives and false negatives, and single-dimensional perception information in urban fire safety monitoring and early warning, this paper proposes a multimodal risk identification algorithm fusing visible light images, infrared thermal images, and timeseries data from smoke and temperature sensors. The algorithm first conducts preprocessing and spatiotemporal alignment for multi-source heterogeneous data, then implements adaptive weighted feature fusion through a channel attention mechanism, and finally optimizes decision results with D-S evidence theory. Experimental results verify that the proposed algorithm significantly surpasses conventional methods in terms of identification accuracy, false positive rate, false negative rate, single-frame inference latency and fire recognition robustness. The algorithm can effectively address the problem of urban fire hidden danger identification and offer technical support for fire monitoring, early warning and emergency disposal.