Point cloud filtering is one of the key steps of airborne LiDAR point data processing. Most traditional methods obtain satisfactory effects just for several specific terrain types. However, ground filtering on the point cloud of complex or mixed terrain types faces a huge challenge. Therefore, a new point cloud filtering method based on grid ground saliency division was proposed. As the point clouds were organized with virtual grid, a ground saliency division based on elevation was performed on the scanning line of point clouds. For different types of grids, different filtering processes were employed to segment point clouds into ground points and non-ground points according to the ground saliency value. Compared with other classical methods, the proposed method avoids the iterative encryption process, and the curved surface was used to fit local terrain in those undulating area, which has better adaptability in complex and mixed terrains with limited increase of computational cost, and generates a set of ground points with high reliability.