ing at the large measurement error and low accuracy problems caused by refraction when conducting underwater measurement based on non-parallel stereovision system, a measurement model of underwater stereovision system is built via refraction light path. Given the relative position relationship between two cameras, a parameter calibration method which is suitable for the measurement model is improved based on Agrawal's algorithm. To verify the feasibility and robustness of the improved method, the underwater calibration experiments are carried out. Experimental results show that, for the parameter of normal vector of waterproof cover, the results of the improved method are closer to the actual results than that of Agrawal's algorithm. The measurement model of underwater stereovision system calibrated in this paper is applied to measure the standard length between the calibration points on the underwater target. The average measurement error is -0.0134 mm, and the maximum error is 0.2073 mm, which is equivalent to the measurement accuracy of stereovision system in air.