VisInt-VHM

To ensure the usability and reliability of the collected data, one Hikvision monitoring camera (iDS-TCV900-AE/25) is deployed at the entrance of Taijia Expressway in Shanxi province in China for image capturing. This camera is installed on the roadside pole with a height of 5.8 meters and uses the infrared flash as the supplementary lighting. The captured images cover two lanes of the expressway, with the resolution being 4096*2160. All images are captured during the period of November 2019 to April 2020.We have collected 10,000 images of front-view hazmat vehicles, considering three aspects of variations. Firstly, there are totally 24 kinds of vehicles with diverse appearances, as shown in Fig.(a). Secondly, these images are captured under different weather and light conditions, which may severely degenerate the quality of resulting images, as shown in Fig.(b). Thirdly, it is very likely that images are captured with dynamic shooting angles, which also poses some difficulty for hazmat marker detection, as shown in Fig.(c). These vehicle images totally consist of 20,023 hazmat markers, including triangle light markers and license plates with different postures and positions, as shown in Fig. (d). All markers in images are annotated with bound-boxes by LabelImg. It should be noted that though there are a wide range of variations about these images, there is only one kind of label for them, since our focus is vehicle hazmat marker detection rather than classification.The download link of the dataset is as follows:Baidu network disk：https://pan.baidu.com/s/16jMFTJJJ6XpjSQBROugvgw key：1234

为保障采集数据的可用性与可靠性，本数据集在中国山西省太嘉高速公路入口处部署了一台海康威视监控摄像头（iDS-TCV900-AE/25）用于图像采集。该摄像头安装于5.8米高的路边杆体，采用红外闪光灯作为补光设备。采集的图像覆盖高速公路两条车道，分辨率为4096*2160，所有图像采集于2019年11月至2020年4月期间。本次共采集10000张危险品运输车辆的正视角图像，涵盖三类变化维度：其一为车辆外观差异，共包含24种外观各异的车型，如图(a)所示；其二为环境条件差异，图像采集于不同天气与光照环境下，此类因素可能严重降低图像质量，如图(b)所示；其三为拍摄角度差异，图像采集时存在动态拍摄角度变化，这也为危险品标识检测带来了一定难度，如图(c)所示。这些车辆图像中共包含20023个危险品标识，涵盖三角警示标识与不同姿态、位置的车牌标识，如图(d)所示。图像内所有标识均通过LabelImg工具完成了边界框标注。需要说明的是，尽管本数据集的图像存在丰富的变化类型，但仅设置单一类别标签，因为本研究的核心目标为危险品车辆标识检测而非车辆分类。本数据集的下载链接如下：百度网盘：https://pan.baidu.com/s/16jMFTJJJ6XpjSQBROugvgw 提取码：1234