In the seam tracking based on structured light vision, strong noise such as arc and spatter generated in the welding process would greatly reduce the visibility of the weld, resulting in tracking failure. In order to solve this problem, a seam tracking algorithm based on the improved kernel correlation filter was proposed to better adapt to the strong noise environment. Firstly, the initial feature points of the weld were obtained. Based on the gray distribution characteristics of the weld laser stripe, the centerline of the weld laser stripe was extracted by Steger algorithm. Then the center line was filtered and derived to obtain the initial feature points of the weld. Finally, the initial feature points of the weld were used as the initial input, and the improved kernel correlation filter was applied to learn and track the weld feature points. The results show that, the average error of the algorithm is 0.305 mm, and the maximum tracking error is 0.479 mm in the case of strong noise interference, which achieves good tracking effect and effectively avoids tracking drift. This study provides a useful reference for high-precision seam tracking.