MLE-YOLO: A lightweight and robust vehicle and pedestrian detector for adverse weather in autonomous driving

Adverse weather poses significant challenges to object detection in autonomous driving, including poor visibility, precipitation interference, and motion blur. Additionally, conventional object detectors often struggle to balance computational efficiency with detection accuracy in such conditions. To address these issues, we propose MLE-YOLO (Multimodal Lightweight Enhanced YOLO), a multimodal fusion framework built upon YOLOv11, optimized for robust vehicle and pedestrian detection in adverse environments. MLE-YOLO integrates four key innovations. The backbone is enhanced with a Multi-Stage Partial Transformer Module (M-SPTM), which combines CNN and Transformer branches to improve detection accuracy in foggy and rainy scenarios while maintaining computational efficiency. The neck adopts a Mixed Aggregation Network (MANet) that leverages depthwise separable convolutions and dynamic cross-layer connections to strengthen multi-scale feature fusion and suppress weather-induced noise. A Lightweight Downsampling Module (LDM) is designed to enhance small-object detection through multi-path feature aggregation, coupled with a compact structure that reduces computational load. Finally, an Efficient Lightweight Detection Head (ELDH) incorporates detail-enhancing convolutions to extract both intensity and gradient features from degraded visual inputs. Extensive experiments on a custom adverse weather dataset demonstrate that MLE-YOLO improves F1 score by 3.6 % and mAP by 3.0 %, while reducing model parameters by 15.8 %, model size by 7.1 %, and FLOPs by 3.2 %. These results validate MLE-YOLO as a lightweight and robust solution for real-time perception in autonomous driving under challenging environmental conditions.