The methods are proposed to improve robustness of detecting the weld seam shaped zigzag line in the case of powerful arc light and light reflex. Firstly, robust estimation methods are proposed to fit the direction of weld seam and the normal vector of workpiece plane, which make the arc welding robot track weld seam with satisfied precision. Secondly, a predictive model for feature points of weld scam is established using Kalman filter to avoid tracking failure. Thirdly, the method based on predictive model is used to decide whether current feature point is a turning point on weld seam shaped zigzag line. Experiments show the effectiveness of the methods