LiDAR-Based Long-Term Mapping in Snow-Covered Environments

TL;DR

This paper introduces a method to improve long-term autonomous driving in snowy environments by detecting and removing snow from maps to maintain accuracy.

Contribution

A novel mapping strategy using deep learning to detect and remove snow, improving map alignment and localization in snow-covered areas.

Findings

The proposed method achieves 78.6% IoU for snow detection.

Map alignment errors are reduced by 12.5% (RMSE) and 15.6% (Chamfer Distance).

The method improves long-term map quality in snow-covered environments.

Abstract

Autonomous driving systems encounter various uncertainties in real-world environments, many of which are difficult to represent in maps. Among them, accumulated snow poses a unique challenge since its shape and volume gradually change over time. If accumulated snow is included in a map, it leads to two main problems. First, during long-term driving, discrepancies between the actual and mapped environments, caused by melting snow, can significantly degrade localization performance. Second, the inclusion of large amounts of accumulated snow in the map can cause registration errors between sessions, thereby hindering accurate map updates. To address these issues, we propose a mapping strategy specifically designed for snow-covered environments. The proposed method first detects and removes accumulated snow using a deep learning-based approach. The resulting snow-free data are then used for map updating, and the ground information occluded by snow is subsequently restored. The effectiveness of the proposed method is validated with data collected in real-world snow-covered environments. Experimental results demonstrate that the proposed method achieves 78.6% IoU for snow detection and reduces map alignment errors by 12.5% (RMSE) and 15.6% (Chamfer Distance) on average, contributing to maintaining map quality and enabling long-term autonomous driving in snow-covered environments.

Figures