This paper has proposed a method to realise weld path identification. Before welding, the area of the weld center is automatically extracted from the image of the weld. Based on the kernel correlation filter algorithm, the area of the weld center can be located in real-time during welding. To reduce noise contamination of the weld target template, the template is updated based on the response map’s peak sidelobe ratio. Specifically, the peak sidelobe ratio is compared with the set threshold to determine template update. A Kalman filtering model of the position of the weld center is established to get the optimal estimation of the weld center and the state amplification method is adopted to process the colored observation noises. The experimental results show that the proposed method can accurately and stably identify the center position of the butted weld and the lapped weld, and the identification errors can be controlled within 0.137 and 0.105 mm, respectively, meeting the requirements of identifying the weld path with high accuracy.