URMDet-SimFire: Trustworthy multimodal detection with uncertainty and reliability modeling for fire and smoke analytics

Le résumé fourni par la source

Trustworthy multimodal perception is essential for safety-critical analytics, where detection systems must contend with noisy visual observations, missing sensor signals, smoke-like interference, and distributional shifts across environments. Vision-only detectors degrade under precisely these conditions, while naively combining additional modalities can amplify corrupted signals. We propose URMDet-SimFire, an uncertainty- and reliability-aware multimodal object detection framework that addresses this fragility by explicitly modeling the trustworthiness of each input source at inference time. We couple a YOLOv8-style visual branch with four non-visual modality encoders for thermal cues, environmental sensors, smoke-diffusion attributes, and textual scene priors. Two trust-modeling components are then introduced: an uncertainty calibration module that adjusts detection confidence using classification evidence and localization variance, and a reliability-guided fusion module that dynamically reweights each modality according to its current noise level, missingness, and cross-modal consistency. Because collecting real industrial fire data at scale is dangerous and costly, we validate the framework in a controllable simulator that reproduces typical visual degradations, sensor corruption, modality dropouts, and occlusion patterns. We evaluate it through main comparisons, ablations of each trust-modeling component, robustness under four corruption types, and efficiency profiling. The results show that naive multimodal late fusion can underperform vision-only detection when modalities are unreliable, that uncertainty calibration alone is insufficient, and that the proposed reliability-guided design recovers detection quality. The framework offers a reproducible simulation-based paradigm for trust-aware multimodal media analytics.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.