Autonomous driving in complex lighting environments by deep learning based object detection

Ziwei Yin

doi:10.1016/j.aej.2026.05.011

Return

Autonomous driving in complex lighting environments by deep learning based object detection

2026-05-20

DOI：10.1016/j.aej.2026.05.011

Original

Save

View PDF

Abstract

En 中文

detection in autonomous driving remains highly susceptible to performance degradation under complex illumination conditions, including low-light environments, overexposure, glare, and spatially uneven brightness. These challenges cause severe feature distortion and reduced detection reliability, especially for small or partially occluded road users. To address this problem, this study proposes LA-DETR (Light-Aware Detection Transformer), an end-to-end real-time detection framework designed specifically for non-ideal lighting scenarios. The framework incorporates a Retinex-inspired Photoadaptive Convolution (PAC) for shallow photometric enhancement and integrates brightness-aware attention mechanisms (BAFM and CBAM) within the encoder to improve illumination robustness and suppress lighting-induced noise. To rigorously evaluate its effectiveness, two structured subsets—BDD100K-Light and BDD100K-Exposure—are constructed through luminance-based filtering and annotation verification to represent typical low-light and overexposure scenarios. Experimental results show that LA-DETR achieves 0.26%–2.51% and 0.19%–1.27% mAP@0.5 improvements on the two subsets, demonstrating superior accuracy and stability compared to existing real-time detectors. In conclusion, LA-DETR provides a practical and illumination-robust solution for autonomous driving perception. Future work will explore lightweight optimization and embedded deployment strategies to enhance applicability in real-world automotive systems.

Keywords:

Intelligent vehicle

Autonomous driving

Artificial intelligence

Object detection

AI Summary

Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.