CrowdHuman Dataset with YOLO Annotations

The dataset contains approximately 12,000 training images, 1,500 validation** images, and 1,500 testing images**, following a standard 70% / 15% / 15% train-validation-test split. The dataset includes a single class, person, and all images are maintained in their original resolutions without resizing to preserve real-world scene characteristics.

The dataset is designed for human detection in crowded environments and contains diverse scenarios with varying illumination, occlusion, motion blur, and dense pedestrian distributions. Some images may include partially visible humans, overlapping persons, blurred regions, and complex backgrounds, making the dataset suitable for robust object detection research.

This cleaned and annotated version of the dataset includes:

Removal of missing, duplicate, and corrupted images. Human bounding box annotations generated using a pre-trained YOLOv8 model. Annotation files provided in standard YOLO .txt format for direct training compatibility. Additional annotations exported in CSV format for analysis and reuse in custom pipelines.

The dataset is beginner-friendly and can be effectively used for:

Human detection projects Crowd analysis research Object detection model training Educational and academic purposes Benchmarking lightweight detection frameworks

该数据集包含约12000张训练集图像、1500张验证集图像以及1500张测试集图像，遵循标准的70%/15%/15%训练-验证-测试划分比例。数据集仅包含单类别：行人（person），所有图像均保留原始分辨率未进行尺寸调整，以保留真实场景的特征。
本数据集专为拥挤环境下的行人检测任务设计，涵盖多样化场景，包含不同光照条件、遮挡情况、运动模糊以及密集行人分布等多种样本。部分图像中可能包含部分可见的行人、相互重叠的人物、模糊区域以及复杂背景，因此该数据集适用于鲁棒性目标检测相关研究。
本数据集为经过清洗与标注的版本，包含以下处理：移除缺失、重复与损坏的图像；使用预训练的YOLOv8模型生成行人边界框标注；标注文件采用标准YOLO .txt格式，可直接适配模型训练；额外导出了CSV格式的标注文件，便于数据分析以及在自定义流程中复用。
该数据集对新手友好，门槛较低，可有效应用于以下场景：行人检测项目、人群分析研究、目标检测模型训练、教学与学术用途，以及轻量级检测框架的基准测试。