Measuring three-dimensional (3D) geometrical parameters of weld pool surface is a key to developing next-generation intelligent welding machines that can mimic a skilled human welder to control welding process through optical observation. Although different techniques have been applied in the past few years, the highly dynamic and specular natures of the molten liquid metal surface in addition to the well-known difficulties associated with the bright arc complicate these approaches. In this paper, a three-dimensional vision-based weld pool surface measurement system for gas tungsten arc (GTA) welding is proposed. In the system, through projecting a low-power continuous structured laser pattern onto the weld pool surface, the specular nature of the surface is taken as an advantage to overcome the difficulties caused by the bright arc. The reflection is intercepted; the image on the imaging plane is captured and analyzed to perform 3-D reconstruction. To successfully derive the pool surface from captured laser pattern reflection, robust image processing algorithms, correspondence simulation method and iterative reconstruction algorithms have been developed. Using these algorithms, dynamic weld pool surfaces have been successfully derived. The authors have also studied the fluctuation of the weld pool surface during gas tungsten arc welding (GTAW) process using proposed approach and conducted a systematic error analysis which can help further improve measurement accuracy. Experimental results verified the effectiveness and accuracy of the proposed measurement system.