Binocular stereo imaging technology has a wide application prospect in underwater archaeology, Marine geological survey, autonomous navigation of underwater robot, underwater biological investigation and other fields. However, when the camera is applied in an underwater environment, the imaging light enters the lens through media of water, window glass and air. Due to the different densities of the three media, the light will be refracted and bend, and the measurement model and calibration method in air will fail, making it difficult to realize underwater 3D reconstruction. In the early studies, the refraction effect was ignored, resulting in low accuracy of underwater 3D measurement. Subsequently, researchers at home and abroad believe that establishing an accurate underwater refraction imaging model is the key to improving the calibration accuracy of binocular stereo measurement system. Therefore, based on the multi-layer plane refraction model, this paper describes the underwater camera imaging process by using the 4D parametric representation method of light, and establishes the underwater binocular stereo vision imaging model. On this basis, this paper proposes the corresponding calibration method and uses the forward projection error as the objective function to optimize the normal vector of parameter interface and the thickness of air medium, and verifies the correctness of the calibration algorithm through MATLAB simulation. The results show that the average error of system parameter calibration simulation analysis is 0.078 pixel, showing that the calibration algorithm is extremely accurate. Then, the system calibration experiment, underwater precision experiment and underwater object 3D reconstruction experiment are carried out. The results show that the average error of the system calibration experiment is less than 0.12 pixel, indicating that the proposed calibration method of underwater binocular stereo vision imaging system has high calibration accuracy of system parameters, which lays a foundation for pixel matching and 3D reconstruction. In the underwater accuracy experiment, the standard deviation of the underwater standard ball rod measurement error is 0.8 mm, the standard deviation of the measurement error is 0.6 mm, and the standard deviation of the ball center distance measurement error is 1.2 mm, indicating that the 3D measurement accuracy based on the calibration algorithm is very high. The point cloud obtained by underwater 3D reconstruction of physical objects is generally dense, without obvious holes, with good and smooth contour and good effect of detail reconstruction, which fully indicates that the underwater stereo vision model and calibration method proposed based on the above theory have good underwater 3D reconstruction effect.