闭路电视事件数据集——跌倒与躺倒检测

1. 简介该数据集是开源合成数据集，专为俯视 CCTV 场景下的坠落检测、姿态估计和事件监测计算机视觉任务设计；区别于标准目标检测数据集，其含边界盒与 17 个 COCO 标准关键点骨架的双重注释，可支撑模型理解人体姿势并区分站立 / 倒下个体；数据集 95% 样本为倒下状态，适配异常检测模型训练，且完全合成无真实人物，规避 GDPR / 隐私风险，兼容 YOLOv8、YOLO11-Pose 模型

1. Introduction This is an open-source synthetic dataset designed for computer vision tasks including fall detection, pose estimation and event monitoring in overhead CCTV scenarios. Unlike standard object detection datasets, it provides dual annotations consisting of both bounding boxes and 17 standard COCO keypoint skeletons, enabling models to comprehend human postures and differentiate between standing and fallen individuals. 95% of its samples depict fallen subjects, which makes it ideal for training anomaly detection models. The dataset is fully synthetic with no real human participants, thus eliminating GDPR-related privacy risks, and is compatible with YOLOv8 and YOLO11-Pose models.

搜集汇总

背景与挑战

背景概述

该数据集是一个专为俯视CCTV场景设计的开源合成数据集，包含边界盒与17个COCO标准关键点骨架的双重注释，适用于坠落检测、姿态估计和异常事件监测。数据集95%样本为倒下状态，完全合成无真实人物，规避了隐私风险，适配异常检测模型训练。