GLEAM: Automated Structural Health Monitoring from Mobile Platforms via LiDAR-Gated Deep Motion Magnification

Vision-based Structural Health Monitoring (SHM) via Video Motion Magnification (VMM) enables dense, non-contact extraction of structural dynamics. However, deploying deep VMM in the field is restricted by two bottlenecks: a strict requirement for static cameras, and extreme sensitivity to dynamic background noise. To enable automated SHM from mobile platforms, we propose a cross-modal pipeline integrating multi-metric SLAM with Swin Transformer-based VMM. First, a SLAM-driven homography actively cancels 6-DoF payload kinematics to establish a rigid “virtual tripod.” Second, a deterministic LiDAR-gated pruning technique dynamically crops the video strictly to the target structure. This eliminates environmental distractors and overcomes the computational limits of Vision Transformers. Structural anomalies are then extracted via variance analysis and projected into a 3D point cloud for spatial localization. Experiments demonstrate this framework successfully decouples micro-vibrations from ego-motion, rejects background noise, and reduces neural network compute by over 92.3%, delivering a robust solution for mobile SHM.