Inspection video-based bolt loosening measurement in railway fastener using bolt tracking and sequence matching

Fastener bolt loosening detection is vital for railway safety. Current three-dimensional vision-based methods depend on costly data collection and computing hardware, while state comparison-dependent two-dimensional (2D) vision-based methods fail to address reliable bolt-state association across inspection cycles. To address these challenges, this article proposes a 2D detection-tracking-matching framework for fastener bolt loosening measurement. First, BoltTransNet detects the bolt reference angle in each video frame to determine its rotational state. Second, the area-based DeepSORT assigns numbers to bolts across frames to construct bolt sequences. A final reference angle for each bolt is then obtained after outlier removal and averaging. Finally, the Needleman–Wunsch algorithm aligns the current bolt state sequence with the initial one to resolve mismatches caused by bolt false and missed detections, and the loosening angle is quantified by comparing the final reference angles. This method demonstrates high accuracy and efficiency, with an average error of 1.67° and a processing speed of 24 ms per-frame. The method can be further integrated with corner-based reference angle optimization and segmented matching to address the challenges such as obscured bolt surfaces and long-sequence matching, highlighting its potential for robust railway inspection.