Abstract: Welding robot is commonly used in the traditional teaching and playback method for getting the welding path. In order to ensure that this work mode can be implemented successfully, manual welding must be adopted to complete the workpiece positioning, which will cause the positioning error and make the actual trajectory deviate from the teaching path. In order to solve this problem, welding path search algorithm based on machine vision is put forward. Pyramid hierarchical template matching based on geometry algorithm is studied, and the offset and rotation are calculated to correct the teaching path. It is proved experimentally that the algorithm has high accuracy and efficiency, which can meet the requirements of real time response ability when the welding robot search trajectory automatically. The industrial application shows that, in the practical welding environment, when the welding workpiece offset in X axis and Y axis direction is less than ±5.5 mm and a rotation angle of 8° in the XOY plane, the algorithm can still search the welding path accurately.