Video Alignment for Predictive Maintenance in Railway Infrastructure

Details

Supervisors

Schaus, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enThis thesis develops an automated method for the temporal synchronization of two rail-track videos recorded along the same track at different times and different speeds, a prerequisite for any downstream change or defect detection. The proposed solution is a coarse-to-fine hybrid pipeline. A global alignment skeleton is first built with Open-End Dynamic Time Warping applied to gradient-orientation histograms, then refined locally by binary-descriptor feature matching (AKAZE, BRISK, ORB) within a strict temporal window, before a dynamic-programming pass enforces monotonicity and a weighted smoothing removes residual jitter. The pipeline runs entirely on CPU. It is evaluated on three real-world sequences using SSIM, LPIPS, and cycle consistency. A key finding is that SSIM and LPIPS saturate on this footage, whereas cycle consistency reliably discriminates alignment quality.