Multimodal Data Fusion for Security Screening with Weighted Confidence Aggregation and Escalation in Edge-cloud Systems

Humanitarian demining requires rapid, accurate and resource-efficient identification of concealed explosive threats. This article presents a multimodal sensor-fusion framework that unites outputs from a visual detector (YOLOv8), two vision language models (ChatGPT--4.1 Vision and Gemini 2.5 Flash Vision) and a custom metal detector into a single integral probability Pfinal. Each modality's raw confidence is calibrated to a comparable probability via temperature or Platt scaling, then fused through learnable non-negative weights or logical rules (OR, AND, k-of-n) to support tunable trade-offs between probability of detection and false-alarm rate. A staged architecture executes preliminary detection on a UAV, confirms candidates at a ground station and escalates cases to the cloud, preserving real-time performance despite limited bandwidth and compute resources.