EMHDD
收藏资源简介:
Compared to daytime, nighttime or other challenging scenarios present lower visibility and larger blind spots. Detection tasks in more challenging scenarios, such as nighttime, still suffer from problems such as low accuracy, which are problems that researchers need to break through. In order to push the field further in the future, we constructed a more comprehensive dataset. Over six months, we collected urban road videos covering various scenarios such as daytime, nighttime, cloudy, and rainy conditions. We reasonably positioned several camera sampling points in both Beijing and Jinan, Shandong Province. These cameras are designed to rotate at scheduled intervals, capturing motorcycles from different angles and directions. Then, we conduct preliminary processing on the collected videos by trimming segments with prolonged absences of motorcycles and removing segments where motorcycles are only present at the edges of the camera lenses. This processing ensures the effectiveness of the data used for model training, mitigating any potential adverse impact on our model. Ultimately, we obtained 265 valid videos. From these videos, we extracted 52,800 images at a rate of 2 frames per second and further annotated these images, creating a more comprehensive dataset, which was named EMHDD (Enhanced Motorcycle Helmet Detection Dataset).
相较于日间场景,夜间及其他复杂场景的可视度更低,盲区范围更广。诸如夜间在内的复杂场景目标检测任务仍存在检测精度偏低等亟待研究者突破的痛点。为推动该领域未来发展,我们构建了一款更为全面的数据集。 历时六个月,我们采集了覆盖日间、夜间、阴天、雨天等多种场景的城市道路视频。我们在北京与山东省济南市合理布设了多个摄像采集点位,这些摄像头按预设周期进行旋转,以获取不同角度、不同方向的摩托车影像。随后,我们对采集到的视频进行预处理:剔除摩托车长时间未出现的片段,以及仅在镜头边缘出现摩托车的片段。该预处理流程保障了模型训练所用数据的有效性,降低了其对模型训练可能带来的负面影响。最终我们得到265条有效视频。我们以每秒2帧的采样速率从这些视频中提取出52800张图像,并对这些图像进行标注,最终构建得到一款更为全面的数据集,命名为EMHDD(Enhanced Motorcycle Helmet Detection Dataset,增强型摩托车头盔检测数据集)。
EMHDD 数据集概述
数据集名称
- 名称:Enhanced Motorcycle Helmet Detection Dataset (EMHDD)
数据集描述
- 目的:为了提高在夜间或其他挑战性场景中摩托车头盔检测任务的准确性,构建了一个更全面的摩托车头盔检测数据集。
- 采集过程:在六个月内,从北京和山东济南的城市道路视频中收集数据,覆盖了白天、夜晚、多云和雨天等多种场景。通过设置多个摄像头采样点,并设计摄像头定时旋转,从不同角度和方向捕捉摩托车。
- 数据处理:对收集的视频进行初步处理,剔除了长时间无摩托车出现的片段和摩托车仅出现在镜头边缘的片段,确保数据的有效性。
- 数据量:最终获得265个有效视频,从中提取了52,800张图像,并以每秒2帧的速率进行标注。
数据集结构
- 结构:数据集分为检测器(Detector)和识别器(Recognizer)两个部分,每个部分根据不同天气条件(白天、夜晚、雨天、多云)进行分类。
数据集示例
- 示例图像:提供了数据集中的示例图像,展示了数据集的实际应用情况。
联系方式
- 联系人:Zhiqiang Liu
- 邮箱:2023211273@student.cup.edu.cn




