For decades much work has been devoted to the research and development of automatic arc welding systems. However, it has remained a challenging problem. Besides the very complex arc welding process itself, the lack of ability to precisely sense the welding process, including the seam geometry and the weld pool, has also prevented the realization of a closed-loop control system for many years, even though a variety of sensors have been developed. Among all the sensor systems, visual sensors have the advantage of receiving visual information and have been drawn more and more attentions. Typical industrial solutions for seam detection such as using laser scanners suer from several limitations. For instance, it must be positioned some distance ahead to the molten pool and may cause problem when dealing with shiny surfaces. Existing techniques for weld pool sensing mostly rely on auxiliary light sources and are dicult to be widely employed in industry. Compared to human welders, existing sensor systems exhibit severe limitations as mentioned above. With the protection of only a welding shield glass, i.e., without any auxiliary illumination (passive), human welders can extract visual information on the weld pool and the nearby seam as the feedback to adjust the welding torch and/or welding parameters. It is an attractive idea from both academic and industrial point of view to develop a vision system without using any auxiliary light sources which can nevertheless extract relevant information. However, interpreting the images captured in a passive way during welding is challenging and may heavily rely on sophisticated image analysis and machine learning techniques. This industrial PhD project has been founded to tackle this problem. For the last three years, we have explored dierent possibilities and thoroughly investigated the development of a passive vision system which is only equipped with a single o-the-shelf CCD camera and optical lters, yet capable of extracting sucient information for the control purpose. From the hardware side, we have studies the selection of proper optical lters to reduce the interference of the extremely strong arc light and controlling the exposure time of the camera on the y to capture dierent images for seam tracking and weld pool sensing. From the software side, we have designed a passive seam detection algorithm based on robust estimation techniques which can detect the seam geometry very close to the weld pool region. For the weld pool boundary tracking, we have proposed three approaches based on deformable models. The rst approach employs in ating balloons and snakes, which are two types of deformable models, to capture the weld pool boundary. The rst approach relies on a special periodical initialization scheme and only work in short-circuit mode. In order to handle other modes of the arc welding process such as spray mode where the strong arc light exists continuously, we have proposed another two approaches in which the initialization does not rely on the short-circuit moment. The essence is that deformable models can be immune to spurious edges caused by strong arc light and/or re ection from the seam by incorporating prior information on regions and boundaries. The main ndings are organized and presented in this dissertation.