The amount of point cloud data is very large, so it is an important research content to reduce the point cloud data reasonably. Aiming at the problems of missing details and containing holes in traditional point cloud reduction algorithm, this paper proposes an adaptive point cloud reduction algorithm based on multi parameter k-means clustering. In this method, k-neighborhood of point cloud is created based on KD tree, curvature and normal features of point cloud data are calculated by surface fitting, and point cloud features and boundaries are detected and preserved by multi parameter mixed feature extraction method; initial cluster center is determined by KD tree index, k-means clustering is conducted, and clustering results are refined according to maximum curvature deviation. This algorithm, curvature sampling method, uniform grid method and random reduction method are applied to different types of point cloud models for experiments. The results show that the proposed algorithm has the lower standard deviation than the latter three methods in complex model, and can retain the detailed feature information of point cloud. In addition, the reduction effect and model integrity of the proposed algorithm are better than those of uniform grid method and curvature sampling method.