Aiming at the hand-in-eye calibration of underwater welding robot visual servo tracking system, the traditional calibration method is difficult to achieve due to the shortcomings of underwater environment, such as light refraction, difficulty in manual target placement, water interference and so on, and the existing calibration method that have the problems of complex process, difficult operation and low positioning accuracy. Therefore, a nonlinear calibration method of underwater active vision based on single reference point is proposed. In this method, firstly, the camera carried by the robot which should be precisely controlled to do more than five times of translation and more than two times of rotation respectively, and the image of a fixed reference point in the scene should be taken, and the camera parameters and hand-eye relationship should be calibrated by using the linear camera model; then, the nonlinear equations should be established according to the homography matrix of the two images, and the accurate camera parameters and hand-eye relationship can be optimized. Finally, the experimental platform of binocular stereo vision system for underwater welding robot is built, and the underwater hand-eye calibration and application experiments are carried out. The experimental results show that the underwater active vision calibration method is easy to select a single reference point, easy to operate, and the algorithm is simple. The positioning accuracy meets the requirements of welding accuracy, which verifies the effectiveness of the algorithm, and provides conditions for underwater welding robot visual servo automatic tracking weld.