To measure the inner diameter of super large cylindrical shells with high accuracy, a visual detection system has been designed. This paper proposes a super large cylindrical shell diameter detection method based on noise classification, bilateral point cloud filtering, and cylinder feature fitting. First, point cloud data of the targeted cylindrical shell are obtained using a laser tracker. The noise in the point cloud data is classified into large-scale noise and small-scale noise. Various filtering algorithms are employed to remove the large-scale noise, whereas an improved bilateral filtering method is used to eliminate the small-scale noise. Second, a cylindrical-feature fitting method is proposed to measure the inner diameter of the cylindrical shell. Finally, real point cloud data of the super large cylindrical shell are collected for experimental verification. The experimental results show that the proposed method fits the cylinder more accurately than traditional tape measurement and roller detection methods. The error between the inner diameter of the cylinder obtained in the fitting experiment and the measured value is 0.21 mm, with a relative error of 0.003%. Compared to traditional measuring methods, the error is reduced by an order of magnitude. The detection time is less than 15 s, improving both detection efficiency and accuracy. The proposed detection method provides technical support for the online detection of the inner diameter of super large cylindrical shells.