Nowadays, the chief development trend in modern welding technology is welding automation and welding intelligence. External angle weld has a certain proportion in mechanical manufacture industries. In the real-time welding process, due to hot deformation and the fixture of workpieces used frequently, torch will detach welding orbit causes deviation, which will affect welding quality. Therefore, elimination weld deviation is the key to the weld automatic tracking system. In this paper, the authors use the self-developed structured light vision sensor system which has significant advantage compared with arc sensors to capture real-time weld images. In the project of VC++6.0 real-time weld image processing, after binaryzation with threshold value seventy, 3*1 median filter, thinning, obtain weld main stripe. Then, using the
extraction algorithm this paper proposed to obtain weld feature points, and compute position of weld. Experiment result verified that the extraction algorithm can locate feature points rapidly and compute the weld deviation accurately.