【摘要】 由于车载系统的硬件资源有限，为了提高分心驾驶检测模型的可部署性和准确率，文中提出一种轻量化检测网络HDSL-YOLO。该模型以YOLOv8n为基础，结合多项优化策略加以改进，主要包括以下方面：首先，引入HGNetV2轻量级主干网络，以有效压缩参数规模，显著提升检测速度与运行效率；其次，引入动态上采样模块（Dysample），增强特征表达能力，特别是在多尺度目标提取上的表现；此外，融合SimAM，进一步强化模型对小目标的感知能力和辨识效果；最后，采用轻量化检测头进一步精简参数。实验结果表明：改进后的HDSL-YOLO算法的mAP@0.5和mAP@0.5:0.95分别达到92.4%和55.4%；与原始YOLOv8n算法相比，改进后的算法不仅检测精度有所提高，而且更加轻量化，实现了双重优化。将HDSLYOLO算法部署到Jetson Nano嵌入式平台，可以更快响应，说明了改进方法的有效性。更多还原

【Abstract】 A lightweight detection network named HDSL-YOLO is designed to avoid the hardware resource constraints of in-vehicle systems and improve the deployability and accuracy rate of distracted driving detection models. On the basis of the YOLOv8n, the model is improved by incorporating multiple optimization strategies, including the following four key aspects: firstly, the HGNetV2 lightweight backbone network is introduced to effectively reduce parameter size while significantly improving detection speed and operational efficiency; secondly, the dynamic upsampling module(Dysample) is integrated to enhance feature representation, particularly in multi-scale object extraction; additionally, the SimAM(simple attention module) is incorporated to further strengthen the model′s perception and recognition capabilities for small objects; and finally, a lightweight detection head is adopted to further streamline parameters. Experiments demonstrate that the mAP@0.5 and mAP@0.5:0.95 of the HDSL-YOLO algorithm is improved by 92.4% and 55.4%, respectively. In comparison with the original YOLOv8n algorithm, the improved algorithm realizes not only an improved detection accuracy rate, but also a further lightweighting, so it achieves dual optimization. Deploying the HDSL-YOLO on the embedded platform Jetson Nano confirms shorter response times, which validates the effectiveness of the proposed improvements.更多还原