LAPRNet: Lightweight Airborne Particle Removal Network for LiDAR Point Clouds

Yanqi Ma, Ziyu Yue, Youwei Wang, Risheng Liu, Zhixun Su, Junjie Cao

Abstract: Autonomous vehicles and mobile robotics usually rely on LiDAR sensors for outdoor environment perception. Airborne particles, such as fog, rain, and snow, introduce undesired measurement points resulting in missing detection and false positives. Hence LiDAR-based perception systems must contend with inclement weather to avoid a significant drop in performance. This paper introduces a lightweight network to infer these undesired measurement points. It mainly consists of three Wide Multi-Level Residual modules (WMLR). WMLR is delicately designed to integrate wide activation, multi-level shortcuts, and shuffle attention seamlessly, to make it an effective and efficient pre-processing tool for subsequent tasks. We also introduce an enhanced LiDAR data representation to boost the performance further. It integrates point cloud spatial distribution with the standard intensity and distance inputs. Thus, two models following the same network architecture but with the standard and enhanced input representation, namely LAPRNet\(_2\) and LAPRNet\(_3\), are proposed. They are trained and tested in controlled and natural weather environments. Experiments on the WADS and Chamber datasets show that they outperform state-of-the-art deep learning and traditional filtering methods by a significant margin. Considering the limited computing resources on edge devices, both LAPRNet\(_2\) and LAPRNet\(_3\) provide an optimal balance between quality and computation to ensure successful deployment. LAPRNet\(_2\) is more efficient, and the parameters and computations of it against WeatherNet are 1.53M vs. 0.39M and 18.4 GFLOPs vs. 4.9 GFLOPs, respectively. The source code will be available on GitHub soon.