Application and Comparison of Deep Learning Methods to Detect Night-Time Road Surface Conditions for Autonomous Vehicles

Zhang, Hongyi; Sehab, Rabia; Azouigui, Sheherazade; Boukhnifer, Moussa · 2022 · Crossref

DOI: 10.3390/electronics11050786

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety challenge of detecting road surface conditions for autonomous vehicles during nighttime, a scenario where existing sensor systems often fail due to low contrast and limited lighting. While deep learning models have demonstrated high accuracy in daytime road condition classification, their performance under night-time conditions remains largely unassessed. The authors aim to develop and compare deep learning architectures capable of accurately classifying dry, wet, and snowy road surfaces at night, thereby enabling vehicles to adjust braking distances and enhance driving safety. The research methodology involves analyzing the optical reflection features of road surfaces under two distinct illumination scenarios: with ambient light (e.g., street lamps) and without ambient light (relying solely on vehicle headlamps). The authors note that wet roads exhibit specular reflection, appearing dark under headlamp-only illumination but bright under ambient light, whereas dry and snowy roads scatter light. Based on these features, a custom database was constructed from public videos, manually labeled, and split into training and validation sets for both illumination conditions. Five deep learning models were evaluated: a custom Convolutional Neural Network (CNN), SqueezeNet, VGG (16 and 19 layers), ResNet50, and DenseNet121. Images were pre-processed using histogram equalization and converted into RGB or HSV color spaces depending on the model requirements. Training was conducted on an RTX 3080 GPU using the Adam optimizer. The results indicate that all models achieved training accuracies near 99%, but validation accuracies varied. DenseNet121 emerged as the superior model, achieving validation accuracies of 94.08% with ambient illumination and 95.46% without ambient illumination. While ResNet50 and other CNN variants performed well (around 90–92% validation accuracy), DenseNet121 offered the best balance of performance and efficiency. It required significantly less storage space (approximately 80 MB) compared to VGG models (>1.5 GB) and only slightly more than SqueezeNet (8.6 MB). Although DenseNet121 had the longest inference time (41 ms per image), this was deemed acceptable for real-time autonomous driving decisions. Confusion matrix analysis revealed that wet conditions were identified with the highest precision, while some dry and snowy surfaces were occasionally misclassified as wet. The significance of this work lies in providing a robust, high-accuracy solution for night-time road surface detection, filling a gap in autonomous vehicle safety systems. By demonstrating that DenseNet121 can achieve over 94% accuracy across diverse lighting conditions with manageable computational requirements, the study supports the practical implementation of such systems in autonomous vehicles. This capability allows for proactive safety adjustments, such as increased braking distances, when encountering slippery surfaces like wet or snowy roads at night, ultimately reducing accident risks associated with adverse weather conditions.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• visibility analysis litigation