Auto-driving is developing rapidly nowadays. Tesla, Nio and other car manufacturers have put their level 2 autonomous driving products on the market. As a kind of important sensors in the car, lidar can directly get the distance and angle to the object. That information is organized into the form called “point cloud”. Point cloud is mainly used to rebuild the 3D scene, which plays a major role in guiding the vehicle. But due to weather and other reasons, there is a large number of noise points in the point cloud data detected by lidar, which will cause the accuracy of 3D reconstruction to decrease, and the structure of the object cannot be fully reproduced. It is dangerous while driving since the vehicle does not have enough information about its environment. So, point cloud denoising is necessary and important.To solve this problem, this paper proposes a three-dimensional point cloud denoising method based on adaptive threshold, which has two stages. According to the Euclidean distance between noise points and non-noise points, this method divides the noise points into two types: far-signal noise points and near-signal noise points. For removing the two types of noise points, threshold adaptive denoising algorithm based on nonlinear function and denoising algorithm based on curvature are used respectively at different stages. At the first stage, the threshold adaptive denoising algorithm based on nonlinear function is to remove the far-signal noise. It uses ordered grids to organize disordered point clouds and calculates the average density of point clouds in the grid. Then, the nonlinear function whose input is the distance from the grid to the lidar is called for calculating the threshold to realize the adaptive adjustment of the denoising density threshold. The points in the grid whose density does not reach its threshold would be seen as noise points. At the second stage, the denoising algorithm based on the median is to remove the near-signal noise. It uses the K-D trees to organize the remaining point cloud data. Then all points in the neighborhood of a point P are sorted by curvature. Finally, the median curvature is calculated, and the point P greater than the median curvature is treated as noise points.To verify the proposed method, a set of experiments was carried out with the dataset from the Stanford 3D scanning repository. The origin data from the Stanford 3D scanning repository was seen as non-noise point clouds. And noise points were added whose amount is about 10% of the original data volume. The experiments showed that this method could effectively remove the noise points in the point cloud, and the denoising accuracy is above 95%.