LiDAR is one of the core sensors in the environmental perception system for autonomous driving. It acquires 3D point cloud data of the surrounding environment by transmitting and receiving laser signals, which provides fundamental information for object detection, obstacle recognition, localization and mapping. However, in real-world autonomous driving scenarios, factors such as rain, snow, fog, dust, direct sunlight, long-distance stray light, specular reflection of objects and inherent hardware errors of the sensor introduce numerous noisy points into the raw point cloud. Noise may lead to distorted object contours, misclassification of ground points, missed detection of dynamic objects and point cloud registration errors, which directly degrade the stability of perception algorithms and pose potential risks to driving safety. This paper sorts out the sources and classification of LiDAR point cloud noise in autonomous driving scenarios, compares and analyzes the principles, applicable conditions, advantages and disadvantages of existing classic point cloud denoising algorithms for such scenarios. Considering the special requirements of autonomous driving for algorithm real-time performance, edge feature preservation and adaptability to dynamic objects, the difficulties faced by current denoising algorithms are summarized, and the development trends of LiDAR point cloud denoising technology oriented to autonomous driving are prospected.