ETW-YOLO: A Lightweight Object Detection Algorithm for Low-Visibility Scenarios in Adverse Weather

Abstract

To address the issues of low detection accuracy, difficult feature extraction and poor robustness in autonomous driving systems under adverse weather (rain, snow, fog and nighttime), this paper proposes a lightweight object detection algorithm based on improved YOLOv11n with three optimizations: embedding the Efficient Multi-Scale Attention (EMA) mechanism into deep backbone layers, replacing the conventional neck network with a novel Task-Aligned Dynamic Decoupled Fusion (C3k2-TADDH) module, and optimizing bounding box regression via the WIoUv3 dynamic non-monotonic focusing loss. Experiments on public adverse weather datasets with intense background noise show that the algorithm achieves 66.48% mAP@0.5, outperforming the baseline by 10.72% with only marginal growth in computation and parameters. It maintains excellent lightweight characteristics and real-time performance, and shows potential for meeting the deployment requirements of vehicle-mounted edge computing devices.