In order to integrate point clouds generated from different points of view quickly and accurately,key points are selected adaptively in this paper by using thresholds according to multiple neighborhoods curvature changing information.We describe the key points′ charac teristics using the coefficients which are got when fitting key points′ neighborhood surfaces with different rad ii.This method has a high degree of recognition and can extract the point cloud surface geometry informati on with a smaller amount of calculation.The initial registration depends on the correspondence selec ted from the minimum distance between corresponding key points′ feature descriptors.After the corre sponding points have been transformed using initial registration result,the distance threshold is set to r emove wrong corresponding relation whose corresponding points′ distance is larger than the threshold.Th e strategy is simple in calculation and has good ability to eliminate wrong matching points.Clustering s orting method is used to optimize the correspondence,which keeps the local corresponding relation that has the maximum similarity between characteristic vectors.It makes correspondence distributed ev enly.At last,the final result is obtained through secondary registration using the optimized correspond ing relation.The experimental results show that the proposed registration method is fast and has good anti-interference ability,and it is suitable for real time 3D measurement applications.