Video-based evaluation of bolt loosening in steel bridges using multi-frame spatiotemporal feature correlation

Bolt loosening in steel bridges poses critical safety risks. This paper proposes a multi-view spatiotemporal framework to assess bolt loosening. First, YOLO detects gusset plates and bolts, with spatiotemporal correlation model extracting region-specific bolt video clips. An enhanced UNet architecture quantifies bolt shadow areas as loosening indicators. To address single-frame feature limitations, phase correlation aligns multi-frame shadow regions, deriving enriched multi-perspective features. Finally, features are normalized within each gusset plate region, enabling probabilistic neural network to determine loosening severity. The framework was validated via a steel bridge semi-physical model under diverse conditions (imaging distances, backgrounds, illumination). Results confirm its reliability in delivering robust evaluations despite environmental variability. • YOLO-driven bolt detection extracted region-specific clips via spatiotemporal modeling. • Multi-scale UNet leveraged shadow areas to quantify loosening. • Phase-correlation operations amplified bolt feature richness. • Normalized gusset plate features enabled accurate PNN evaluation.