The conventional feature preserving point cloud simplification method needs to calculate the differential information of all point clouds, but there is a certain deviation in the results by direct calculation with the high density or noise-containing point cloud, resulting in poor effect of point cloud simplification. We present a point cloud simplification method based on grid dynamic partitioning. Firstly, the model is divided into space grids in which the interference points are eliminated with the random sample consensus method. Secondly, the flatness value of grid is calculated by using the least squares method in remaining points, judging whether the grid needs to be subdivided according to the flatness value. Thirdly, the flat areas are achieved and pressed into large spacing grid, and the features-rich areas are divided into small grids as well. For the points in small grids, Gaussian function is introduced to reduce the weight of distant points for recognition features, and the feature points are identified by integration of the surface variation and neighborhood normal vector angle information and then retained. Points in the large grid are sampled at different sampling rates according to the grid spacing. Comparative experiments are carried out with the random sampling method, grid method, curvature method and the proposed method. It is shown that this method can maintain the fine features of model and avoid the appearance of holes, and the maximum deviation of the simplified model is 1.502 mm, much smaller than those of the other three methods. Moreover, as the noise intensity increases, the simplification error of this method is small and gentle. Under the noise condition of 35 dB, the average deviation is only 40% of those of random sampling method and grid method, as well as 50% of that of the curvature method.