面向智能车辆分类的YOLOv8m改进：α缩放梯度归一化Sigmoid激活函数

摘要

车辆分类在智能交通系统（ITS）与现代交通管理的发展中至关重要，实时准确地检测与识别车辆是维持道路通行效率与安全的关键。本文通过改进YOLOv8m模型的激活函数，使其在多样的交通与环境条件下获得更高的准确率与更快的响应。本研究把Mish与Swish两种替代激活函数集成到YOLOv8m结构中，并与模型默认的Sigmoid线性单元（SiLU）进行对比测试。训练与评估使用了在不同光照与天气条件下采集的车辆数据集。实验结果表明，改进后的激活设计能带来更好的模型收敛性、更强的泛化能力与显著提升的检测性能，准确率最高比标准YOLOv8m提高5.4%，mAP最高提高6.6%。总体而言，上述结果证实：对激活行为进行精细调节，可使深度学习模型在真实智能交通环境下的车辆分类任务中更具适应性与可靠性。

Abstract

Vehicle classification plays a vital part in the development of intelligent transportation systems (ITS) and modern traffic management, where the ability to detect and identify vehicles accurately in real time is essential for maintaining road efficiency and safety. This paper presents an enhancement to the YOLOv8m model by refining its activation function to achieve higher accuracy and faster response in diverse traffic and environmental situations. In this study, two alternative activation functions—Mish and Swish—were integrated into the YOLOv8m structure and tested against the model's default sigmoid linear unit (SiLU). Training and evaluation were carried out using a comprehensive dataset of vehicles captured under different lighting and weather conditions. The experimental findings show that the modified activation design leads to better model convergence, improved generalization, and a noticeable boost in detection performance, recording up to 5.4% higher accuracy and 6.6% better mAP scores than the standard YOLOv8m. Overall, the results confirm that fine-tuning activation behavior can make deep learning models more adaptive and reliable for vehicle classification tasks in real-world intelligent transportation environments.