Intersection tracking based on line segment matching for automated structural seismic response monitoring from occluded surveillance videos

To address the challenge of random occlusions from earthquake-induced object motion in surveillance videos, this study introduces a novel intersection tracking method based on line segment matching. The method employs few-shot transfer learning (TL) to adapt the model for high-precision line segment detection and descriptor extraction, especially in specific monitoring environments with limited training data. Inspired by the unique sequencing of DNA base pairs, this approach achieves robust matching of structural line segments by analyzing the sequence information of their descriptors. The intersections of these matched segments then serve as persistent tracking points, effectively overcoming tracking failure caused by occlusion. Experiments demonstrate that this method significantly outperforms mainstream methods. Even when the line segment occlusion rate reaches 40%, the coefficient of determination (R²) value remains at 0.91. Furthermore, the method maintains high precision in moderately dim scenes and under low-resolution conditions. This approach can automatically and efficiently transform surveillance videos into quantifiable structural seismic response data, offering strong support for earthquake analysis based on surveillance cameras.