In wheel set online detection based on structured light projection image technique, the wheel set light-section curves acquired by CCD are heavily distorted. The camera calibration method and calibration precision will directly affect the wheel set online detection precision. According to the nonlinear distortion imaging model, Tsai non-coplanar calibration algorithm based on a moving planar calibration plate was proposed. According to the radial alignment constraint and first-order radial distortion formula, the least square method and Levenberg-Marquardt optimization algorithm were used to calibrate the internal and external parameters of camera. The calibration plate and its three installation location were designed, on the basis of the requirements of on-site installation position and the key point regional distribution of the light-section curves. The extraction of image feature points and the calibration of camera parameters were realized. The experimental results show that the calibration precision of Tsai non-coplanar calibration algorithm was less than 0.1 mm. The calibration precision meets the demands of wheel set online detection.