GAN Inversion Based Point Clouds Denoising in Foggy Scenarios for Autonomous Driving

Currently, point clouds captured in adverse weather conditions are often subjected to noise, which significantly impacts the reliability of autonomous driving perception systems. To address this issue, we propose a method for denoising point clouds under adverse weather conditions by utilizing a pre-trained generative model to establish a GAN inversion network. To simulate foggy driving scenarios, we construct the Foggy KITTI dataset and pre-train the GAN to capture rich semantic information. The experimental results demonstrate that the proposed method outperforms other denoising methods, indicating its effectiveness in enhancing the quality of point clouds. Additionally, our method exhibits good generalization ability on real-world datasets, indicating its potential as a preprocessing component for other point cloud processing tasks.