YOLO-DH: Robust Object Detection for Autonomous Vehicles in Adverse Weather

Although autonomous driving technology has advanced rapidly in recent years, the intrinsic complexities of adverse weather conditions pose significant challenges to achieving high object detection accuracy, despite extensive research on perception systems in autonomous vehicles. Accordingly, this study proposes a YOLO-based network that integrates the DHNet residual network for image dehazing with wavelet-based channel attention modules, specifically tailored for target detection in autonomous driving under adverse weather conditions. The proposed network introduces a novel channel-value attention mechanism and sequentially applies wavelet-based soft-thresholding denoising techniques, thereby improving its sensitivity to discriminative channel information. We further introduce the first DHNet Dehazing Attention Module, which sequentially combines the DHNet residual network with the MixDehazeNet hybrid structural dehazing network. This integration synergistically improves the network’s image dehazing performance. To address dataset class imbalance, this study incorporates Adaptive Threshold Focal Loss (ATFL), which markedly improves training efficiency and model robustness. This optimization enhances the model’s generalization ability for target detection and classification tasks under challenging weather conditions. Experimental evaluations, including ablation studies and comparative analyses, demonstrate that the proposed method achieves substantial improvements in accuracy across three benchmark datasets.