MonoGuard: towards infrastructure-aware oversized vehicle detection via dynamic 3D metrology with monocular surveillance cameras

Oversized vehicles pose a significant threat to public safety and urban infrastructure, leading to catastrophic bridge failures and traffic disruptions worldwide. However, existing detection systems often lack the automation, real-time capability, and cost-effectiveness required for widespread deployment. To address this gap, this study introduces MonoGuard, a deep learning-enhanced framework for real-time oversized vehicle detection (OSVD). By leveraging ubiquitous monocular surveillance cameras, Monoguard is designed to guard against infrastructure damage and enhance traffic safety. Monoguard integrates three key innovations: a semi-automatic calibration workflow using Segment Anything Model 2 (SAM2) and multi-frame fusion for robust vanishing point estimation; the lightweight Context-Guided Attention Segmentation-You Only Look Once (CGAS-YOLO) model for efficient vehicle segmentation; and a rule-based adaptive 3D bounding box pipeline that dynamically adjusts to camera-object geometry. Extensive evaluations across 36 UAV-simulated scenarios demonstrate MonoGuard’s high performance. It achieves a remarkable height estimation accuracy rate of 96.98% for cars and 95.50% for trucks, while maintaining a real-time throughput of 46.5 FPS on an RTX 4080 Laptop. By repurposing existing surveillance infrastructure, MonoGuard provides a scalable and economical solution for smart cities, enabling early warnings to prevent collisions, protect infrastructure, and safeguard lives and property.