Vision-based seam tracking is a crucial technology widely used in robotic welding. However, during the welding process, intense noise from the arc and spatter leads to large welding seam tracking errors. To solve this problem, this paper proposes a seam tracking method that integrates a feature-fused Kernelized Correlation Filters (FF-KCF) with a Generative Adversarial Network. Firstly, FF-KCF fuses histogram of oriented gradients (HOG) and scale invariant local ternary pattern (SILTP) features to overcome the limitations of individual features and capture a comprehensive feature representation of the weld seam. Secondly, a feature-supervised Conditional Generative Adversarial network is introduced to repair laser stripe images affected by intense noise, addressing model drift issues in noisy environments and long sequence tracking. Model parameter updates and image repair frequencies are controlled by the Peak-to-Sidelobe Ratio (PSR) to enhance tracking efficiency. Finally, a direction clustering least squares fitting algorithm (DC-LS) is proposed to identify feature points in the initial frame image, solving the problem of manually specifying the tracking object in the initial frame required by the FF-KCF. The proposed methodology is validated on the collected dataset. The experimental results show that the tracking speed can reach up to 17.3 FPS, and the average error is kept within 2.2 pixels. Compared to state-of-the-art methods, the proposed method demonstrates superior performance in terms of accuracy and stability, meeting the requirements for precision and real-time seam tracking.