In order to solve the problem of low efficiency and low precision for traditional “teaching method” robot welding, the technology of seam image recognition and tracking based on visual sensing technology was studied in this paper. Median filter and threshold segmentation were used to preprocess the image, and dynamic ROI extraction and least square method were used to extract the feature points of weld seam to solve the influence of arc, spatter and other factors on the image. At the same time, a laser seam tracking system was designed, and the welding torch attitude was taken as an index to analyze the tracking effect of the seam. The results indicate that the designed visual sensing system could effectively extract the welding image features and has high recognition accuracy. The seam tracking system can correct the position of welding torch in real time. The height error of welding torch is always within 0.2 mm, and the angle recognition error of welding torch is not more than 0.4°. This shows that the system has strong antiinterference ability, high welding precision and good realtime performance.