<p>This paper presents the design and experimental validation of an automated weld seam tracking system integrating a laser vision sensor with a multi-axis gantry-based welding robot. In automated welding, accurately detecting feature points of the weld groove from laser stripe images is crucial yet challenging owing to noise from spatter, fumes, and specular reflections. To address this challenge, a robust noise filtering algorithm combined with the YOLOv8-pose deep learning model is introduced. The YOLOv8-pose model provides a high inference speed and detects keypoints and their interconnected skeletal structures analogous to joints. In the proposed approach, the laser stripe generated by the laser vision sensor is considered a skeletal structure, and critical feature points necessary for seam tracking are defined as the corresponding keypoints. By specifically training the model to recognize the unique geometric characteristics of the laser stripe, the weld seam is clearly distinguished from surrounding noise, enabling precise and real-time seam tracking. Furthermore, a supporting algorithm mitigates residual noise induced by fumes or tack welds that cannot be fully filtered by YOLOv8-pose by detecting abnormal geometric variations in the laser stripe. Experimental results demonstrate that the proposed approach achieves a high keypoint detection accuracy with a mean absolute error of 0.128&#xa0;mm, substantially enhancing productivity and ensuring consistent weld quality in industrial shipbuilding environments.</p>