An improved point cloud denoising method in adverse weather conditions based on PP-LiteSeg network

Abstract

Reliable point cloud data (PCD) generated by LiDAR are crucial to perceiving surroundings when autonomous driving systems are a concern. However, adverse weather conditions can impact the detection range of LiDAR, resulting in a significant amount of noisy data that substantially deteriorates the quality of PCD. Point cloud denoising algorithms used for challenging weather conditions suffer from poor accuracy and slow inferences. The manuscript proposes a Series Attention Fusion Denoised Network (SAFDN) based on a semantic segmentation model in real-time, called PP-LiteSeg. The proposed approach provides two key components to the model. The insufficient feature extraction issue in the general-purpose segmentation models is first addressed when dealing with objects with more noise, so the WeatherBlock module is introduced to replace the original layer used for feature extraction. Hence, this module employs dilated convolutions to enhance the receptive field and extract multi-scale features by combining various convolutional kernels. The Series Attention Fusion Module (SAFM) is presented as the second component of the model to tackle the problem of low segmentation accuracy in rainy and foggy weather conditions. The SAFM sequentially applies channel and spatial attention mechanisms to enhance the model’s sensitivity to crucial features. Furthermore, weighted feature fusion is employed to enhance the model’s efficiency in integrating low-level and high-level feature information configurations. Experimental evaluations were conducted on the publicly available DENSE dataset. The results demonstrate that the improved model achieved an 11.1% increase in denoising accuracy measured by MIOU and an inference speed of 205.06 FPS when compared to the PP-LiteSeg model. As a result, the noise recognition accuracy and denoising capability in real-time are enhanced.