3D LiDAR is an important part of the environment perception in the fields of unmanned driving and robots. There is a large number of noise data such as outliers in the point cloud data obtained by scanning the surrounding environment. Aiming at the problem of inaccurate filtering of noise points, this paper innovatively proposes an improved density clustering (DBSCAN) algorithm, which uses a grid network to find the grid with the highest density of point clouds, selects the initial points in it, and adjusts the clustering radius and parameters adaptively, and finally complete the clustering filtering. The experimental results show that the algorithm can effectively filter out outliers while retaining the original point cloud features, and the effect is better than other traditional filtering algorithms when the filtering time is basically unchanged.