Low-latency seam localization for narrow-gap laser welding on industrial edge hardware

Vision-based seam localization in narrow-gap laser welding is challenging because the small joint width and tight positioning tolerances require highly accurate trajectory estimation. Scratches, stamping marks, and varying illumination can generate seam-like image features that cause conventional gradient-based methods to lose the actual joint. This study presents a low-latency hybrid pipeline that combines lightweight semantic segmentation with stateful sub-pixel trajectory extraction for seam localization on industrial edge hardware. The pipeline was evaluated via chronological offline replay of images acquired during welding; no closed-loop control was implemented. To address these artifacts while limiting algorithmic latency for future online integration, the pipeline decouples semantic scene understanding from geometric feature extraction. A lightweight STDC-Net first isolates the joint topology from interfering surface structures. Subsequently, a stateful, moment-based orthogonal tracking algorithm converts the resulting probability map into continuous trajectory coordinates. Offline evaluation on structural-steel joints yielded a root mean square tracking error of 22.2 μ m , with all evaluated trajectory points remaining below an absolute error of 100 μ m . During replay, the measured algorithmic processing time remained within the 5.00 ms inter-frame budget at 200 frames / s without incurring an increasing backlog. The results demonstrate accurate and computationally efficient seam localization under the evaluated conditions.