Multi-Echo Denoising in Adverse Weather

TL;DR

This paper tackles LiDAR degradation in adverse weather by introducing multi-echo denoising, which selects object-relevant echoes while discarding weather-induced noise. It presents SMEDNet, a self-supervised framework that leverages a multi-echo neighbor encoder and two learners (coordinate and correlation) with a blind-spot strategy, complemented by a characteristics similarity regularization to speed up convergence. The method achieves state-of-the-art performance on semi-synthetic SnowyKITTI and proves effective qualitatively on real STF data, while maintaining efficient runtime and reduced parameter count compared to prior self-supervised approaches. The approach promises safer autonomous driving in adverse weather by enabling more reliable, single-echo-like point clouds through multi-echo substitutions. Significant practical impact lies in enhancing perception robustness without requiring dense labeled data or heavy computational overhead.

Abstract

Adverse weather can cause noise to light detection and ranging (LiDAR) data. This is a problem since it is used in many outdoor applications, e.g. object detection and mapping. We propose the task of multi-echo denoising, where the goal is to pick the echo that represents the objects of interest and discard other echoes. Thus, the idea is to pick points from alternative echoes that are not available in standard strongest echo point clouds due to the noise. In an intuitive sense, we are trying to see through the adverse weather. To achieve this goal, we propose a novel self-supervised deep learning method and the characteristics similarity regularization method to boost its performance. Based on extensive experiments on a semi-synthetic dataset, our method achieves superior performance compared to the state-of-the-art in self-supervised adverse weather denoising (23% improvement). Moreover, the experiments with a real multi-echo adverse weather dataset prove the efficacy of multi-echo denoising. Our work enables more reliable point cloud acquisition in adverse weather and thus promises safer autonomous driving and driving assistance systems in such conditions. The code is available at https://github.com/alvariseppanen/SMEDNet

Figures (4)

• Figure 1: In the concept of multi-echo denoising, multiple echoes are acquired for a single emitted pulse and the echo caused by the object of interest is picked. Generally in single-echo approaches, only the strongest echo (red) is available. Thus, crucial information about the object of interest can be lost. I and R denote intensity and range, respectively.
• Figure 2: In the proposed multi-echo neighbor encoder, KNN sets are computed using multi-echo query and KNN reference, we use the strongest echo point cloud as it is de facto in single-echo approaches. Two example queries illustrate how KNN sets look. "Exclude self -- Include neighbors" simply discards KNN queries and keeps the values. From this, the coordinate learner predicts the queries. The point cloud on the right illustrates the final output of our method, where substitutes are points from alternative echoes.
• Figure 3: The proposed self-supervised multi-echo denoising architecture (SMEDNet). White modules are used for training only.
• Figure 4: