Abstract: As a traditional welding method, arc welding is one of the most important ways to achieve the connection of metal materials in industrial production. With the rapid development of robots, arc welding has also shifted from manual operation to automation. At present, the commonly used “teaching demonstration” mode can no longer meet the requirements of intelligent welding. Visual sensing, with its advantages such as non-contact, fast response speed and economic benefits, is promoting automated welding towards intelligence. This paper starts from the demands of intelligent welding at different stages and reviews the development of visual sensing in three aspects, initial weld point positioning, weld seam identification and tracking, and molten pool monitoring. Initial weld joint positioning is mainly achieved through active vision, with the focus on optimizing vision algorithm to adapt to different welding scenarios. Weld seam identification and tracking are divided into three stages, image acquisition, image processing and feedback control, which are mainly achieved through point cloud reconstruction and segmented processing methods. The state of molten pool is related to the stability of welding process and the quality of weld seam. In intelligent welding, morphology and temperature of molten pool are usually monitored as the basis for feedback regulation of welding parameters. Finally, the future development of this field is prospected.