Aiming at the problem of multi-scale feature detection of aircraft part geometric models， a robust multi-scale feature point detection algorithm is proposed. The algorithm first designs a median filter algorithm to obtain the precise normals of unstructured point clouds. Then the local neighborhood fluctuations of each point is calculated based on the calculated normals， and the initial feature points is extractd. Finally， for the data redundancy of the initial feature points， a shrinkage optimization strategy is proposed to calculate the final feature data points. Experimental results show that compared with the traditional point cloud feature point detection method， the algorithm proposed in this paper has significantly improved detection accuracy. Secondly， this method is able to extract precise feature points， even when the input is corrupted by heavy noise.