On the Importance of Quantifying Visibility for Autonomous Vehicles under Extreme Precipitation

TL;DR

This paper introduces a real-time metric for quantifying lidar visibility loss during extreme weather, evaluates its impact on localization accuracy, and highlights the need for datasets focusing on extreme snowfall events.

Contribution

It proposes a novel meteorology-inspired metric for lidar visibility loss, correlates it with localization performance, and analyzes the robustness of ICP under snow conditions.

Findings

ICP algorithm is robust against snowfalls

Snow gusts significantly impair localization accuracy

Extreme weather events require specialized datasets

Abstract

In the context of autonomous driving, vehicles are inherently bound to encounter more extreme weather during which public safety must be ensured. As climate is quickly changing, the frequency of heavy snowstorms is expected to increase and become a major threat to safe navigation. While there is much literature aiming to improve navigation resiliency to winter conditions, there is a lack of standard metrics to quantify the loss of visibility of lidar sensors related to precipitation. This chapter proposes a novel metric to quantify the lidar visibility loss in real time, relying on the notion of visibility from the meteorology research field. We evaluate this metric on the Canadian Adverse Driving Conditions (CADC) dataset, correlate it with the performance of a state-of-the-art lidar-based localization algorithm, and evaluate the benefit of filtering point clouds before the localization process. We show that the Iterative Closest Point (ICP) algorithm is surprisingly robust against snowfalls, but abrupt events, such as snow gusts, can greatly hinder its accuracy. We discuss such events and demonstrate the need for better datasets focusing on these extreme events to quantify their effect.