Abstract: In order to enable the loading robot to accurately acquire the position and attitude information of crankshaft parts, a visual measurement method for crankshaft parts based on 3D point cloud segmentation is proposed. Aiming at the problem of under-segmentation or over-segmentation that may occur when segmenting point cloud data with sharp edges or highly discontinuous surfaces, a normal differential segmentation algorithm combining with area growth is proposed to determine the edge of the area growth with curvature as the judgment threshold and segment the morphological features of the complex shaft parts, so as to solve the problem of the traditional segmentation algorithms which are difficult to segment the continuous surfaces effectively. Then, for the problem of noise and outliers in the point cloud data, it is proposed to use the overall least squares method based on singular value decomposition to obtain the key feature dimensions of complex shaft parts by solving the minimum value of the error equation. Experiments show that the method in this paper can correctly segment highly discontinuous surface features, the average error of segmentation is about 0.095 mm, and the measured crankshaft shaft diameter is more accurate than that of the traditional algorithm.