AW-FRVP: Unsupervised Efficient Point Cloud Denoising for IoT Participants in Adverse Weather

We proposed an unsupervised pipeline AW-FRVP for denoising of point clouds in adverse weather to alleviate the noise points’ impact on perception. Specifically speaking, AW-FRVP consists of two stages: firstly, AWDenoiseNet was adopted to gain efficient pseudo labels by converting the point cloud into frequency domain with depth and intensity features. Secondly, the semantic segmentation network FRVPNet was trained based on the generated pseudo labels and tested for binary classification of noise and environmental points. In the first stage, the KNN-FLATTEN and Mamba-RFE feature extractor were equipped to alleviate the many-to-one issue while converting the point cloud to a range perspective and strengthening the key feature extraction in the frequency domain. In the second stage, point features corresponding to different perspectives were transformed and fused, and two modules named FPFusion-with-CSMA, as well as Triple-Points-Fusion-head, were adopted to gain point-wise features with rich information for subsequent binary classification. Amount of experiments show that the AWDenoiseNet can gain real-time and efficient point cloud denoising with precision of 93.47% on WADS and recall of 94.70% on Weather-KITTI (Rain Scenes). Without annotation, the AWDenoiseNet can gain the SOTA performance among the unsupervised denoising methods. What’s more, the training process of FRVPNet based on pseudo labels generated by AWDenoiseNet can gain sound denoising performance. Both AWDenoiseNet and FRVPNet can achieve real-time denoising with frequencies of about 165 FPS (Frame Per Second) and 19 FPS. In summary, our two-stage unsupervised denoising pipeline can achieve real-time SOTA label-free denoising with a slight gap to those fully-supervised methods (3.65% in precision on WADS, and 5.71% in recall on rain scenes of Weather-KITTI) and our method can significantly reduce the interference of noise in point clouds on perception algorithms for internet of traffic participants.