Point cloud reduction is an important step of reverse engineering. The quality of reduction is directly related to the efficiency of subsequent surface reconstruction. This paper analysed several commonly used methods of point cloud reduction and proposed an improved method against the lacks of existing methods. This method uses the PCA principal component analysis to fit tangent plane with K-Nearest Neighbour points. It calculates the distance between the point and the plane as the basis for initial reduction. Then it uses uniform grid for resampling process to retain some feature points. Through initial reduction and quadratic reduction to finish point cloud reduction. At last, it validates the effectiveness of the method with experiments.