FHWA Vehicle Classification Dataset

This study introduces a significant advancement in vehicle classification, addressing the challenge of limited annotated datasets compliant with Federal Highway Administration (FHWA) guidelines. We present a novel benchmark dataset meticulously curated from various sources to capture variations in time, resolution, camera position, and weather conditions. With a total of 17,174 annotated instances across 7,980 frames, this dataset offers a remarkable granularity for vehicle classification, making this study the first of its kind, to the best of the author’s knowledge. Our research also aims at detecting vehicle subcategories within the FHWA's classification scheme. Acknowledging the visual complexity in distinguishing vehicles with similar appearances but differing weights according to FHWA's criteria, we propose a refined classification system. This system categorizes vehicles into six subcategories based on axle count and spacing, facilitating easier and more precise classification. In addition, we proposed an improved YOLOv5 model that incorporates the Convolutional Block Attention Module (CBAM). The proposed model achieved performance scores of 0.981, 0.965, and 0.985 for precision, recall, and mAP_@50, respectively. As a result, the proposed model outperformed all previous YOLO iterations on the experimental test dataset.  The addition of CBAM improves feature representation by focusing on important elements while ignoring irrelevant ones. The results show that the YOLOv5-CBAM integration is more precise and faster.

本研究在车辆分类领域取得了重要进展，解决了符合美国联邦公路管理局（Federal Highway Administration，FHWA）规范的带标注数据集稀缺这一难题。我们构建了一个全新的基准数据集，该数据集从多源数据中精心筛选整理而来，能够覆盖时间、分辨率、拍摄机位以及天气条件等多种变化场景。该数据集共涵盖7980帧图像，包含17174个带标注的目标实例，为车辆分类任务提供了极高的细粒度标注，据笔者所知，本研究尚属同类首次尝试。本研究同时旨在基于美国联邦公路管理局的分类体系，对车辆子类别进行检测。鉴于根据美国联邦公路管理局的判定标准，外观相似但重量不同的车辆难以通过视觉区分，我们提出了一套优化后的分类体系。该体系基于车轴数量与间距将车辆划分为六个子类别，可实现更简便且精准的分类。此外，我们提出了一种融入卷积块注意力模块（Convolutional Block Attention Module，CBAM）的改进型YOLOv5模型。所提模型在精确率、召回率以及mAP@50指标上分别达到了0.981、0.965与0.985的性能得分。实验结果表明，该模型在测试数据集上的表现优于所有前代YOLO系列模型。卷积块注意力模块的加入通过聚焦关键特征、忽略无关信息，优化了特征提取与表示能力。实验结果证实，集成了CBAM的YOLOv5模型兼具更高精度与更快速度。