Weather and Meteorological Optical Range Classification for Autonomous Driving

Abstract

Weather and meteorological optical range (MOR) perception is crucial for smooth and safe autonomous driving (AD). This article introduces two deep learning-based architectures, employing early and intermediate sensor fusion and multi-task strategies, designed for concurrent weather and MOR classification in AD. Extensive experiments employing the publicly available FogChamber dataset demonstrate that the proposed early fusion architecture, characterized by its lightweight design and simplicity, achieves an accuracy of 98.88% in weather classification and 89.77% in MOR classification, with a competitive memory allocation of 5.33 megabytes (MB) and an inference time of 2.50 milliseconds (ms). In contrast, the proposed intermediate fusion architecture prioritizes performance, achieving higher accuracies of 99.38% in weather classification and 91.88% in MOR classification. However, it requires a more substantial memory allocation of 54.06 MB and exhibits a longer inference time of 15.55 ms. Compared to other state-of-the-art architectures, the proposed methods present a competitive balance between accuracy performance, inference time, and memory allocation, which are crucial parameters for enabling autonomous driving.