When teaching welding robot performs the continuous welding process of automobile thin plate parts,due to the problems of clamping error and thermal deformation,the error between actual trajectory and teaching trajectory of the welding seam is large.In order to improve the welding quality,a laser vision weld detection and tracking system was constructed based on a welding robot,and a weld seam tracking algorithm based on target estimation guideline was proposed to track the three-dimensional position change of the weld center.The initial frame weld feature points were extracted by using traditional image processing method,and the improved siamese network was used to track and extract the welding seam feature point under strong interference.The three-dimensional coordinates of the welding seam center feature points in the robot base coordinate system were obtained through coordinate conversion.The results show that the by using the algorithm,the welding seam feature point can be accurately extracted and tracked,with an average error of 0.48 mm and an average frame rate of 90 f/s,which is better than traditional image processing methods and methods based on correlation filtering,and fast and accurate tracking can be achieved.