A Scene-Aware Confidence-Guided YOLO Framework for Robust Weld Seam Tracking Under Intense Arc Interference.

Accurate weld seam tracking under severe arc interference remains a major challenge for intelligent robotic welding systems because radiometric saturation and transient visual degradation significantly reduce the reliability of vision-based measurements. Existing YOLO-based tracking methods generally use detector outputs directly for state estimation, assuming that the detection confidence adequately reflects localization reliability. However, this assumption often breaks down under strong arc interference, leading to unstable tracking and occasional tracking failure. To address this problem, this paper proposes a scene-aware confidence-guided YOLO (SACG-YOLO) framework for robust weld seam tracking. The proposed method introduces a scene-aware confidence estimator that evaluates the observability of the current welding scene by jointly considering structural consistency and photometric variation. The estimated scene confidence is fused with the detector confidence to assess measurement reliability, which is subsequently used to adaptively regulate the Kalman filter update process. As a result, reliable observations are incorporated into state estimation, while unreliable measurements are rejected in favor of motion prediction, thereby improving tracking continuity during severe visual degradation. Experiments on industrial welding image sequences demonstrate that the proposed framework consistently outperforms conventional YOLO-based tracking methods. Under stable and weak arc interference, the average tracking error is reduced from 0.492 mm to 0.155 mm, while the maximum error decreases from 0.829 mm to 0.354 mm. Under severe arc interference, where standalone YOLO tracking fails, SACG-YOLO maintains continuous weld seam localization with a maximum tracking error of 0.406 mm. These results demonstrate that explicitly modeling measurement reliability through scene-aware confidence estimation significantly improves the accuracy, robustness, and continuity of weld seam tracking in complex industrial welding environments.