In order to solve the problems of unintelligent detection of weld points, low detection accuracy and poor detection robustness of the robot system in the automated welding process of the internal anti-surge plate of aluminum alloy tanks of large pressure vessels, a multi-feature fusion 3D weld point real-time detection method for structural light visionguided robots is designed. First, the visual features are extracted from the workpiece to derive multiple regions of interest, then the multiple features are fused to obtain the 2D key points, and finally the 3D pre-weld points are determined by fast and unconstrained system calibration. Industrial field experiments show that the maximum error of 3D welding point extraction in camera coordinate system is 0.196 mm, the average error is 0.099 mm, and the average detection time is 0.09 s. The welding point detection is accurate, fast and intelligent, which can meet the industrial robot welding path planning and automatic welding tasks.