PHEVA

该数据集名为PHEVA，是一个保护隐私的人本伦理视频异常检测数据集，通过移除像素信息并仅提供去标识化的人类标注来保护个人可识别信息。PHEVA包含了一个庞大的测试集，拥有超过70万帧（约6.5小时）的数据，其规模是第二大数据集的四倍，正常帧与异常帧的比例为70/30。此外，该数据集还包含了超过五百万帧（约46小时）的标注姿态序列，其任务是进行基于姿态的异常检测。

Named PHEVA, this is a privacy-preserving human-centric ethical video anomaly detection dataset that protects personally identifiable information (PII) by removing pixel-level information and only providing de-identified human annotations. PHEVA includes a large-scale test set with over 700,000 frames (approximately 6.5 hours) of data, whose scale is four times that of the second-largest dataset, and the ratio of normal frames to anomalous frames is 70:30. In addition, the dataset also contains over 5 million frames (approximately 46 hours) of annotated pose sequences for pose-based anomaly detection tasks.

搜集汇总

背景与挑战

背景概述

HuVAD是一个专注于视频异常检测的隐私保护数据集，通过去标识化的人体姿态标注（而非原始像素）来确保隐私安全，适用于以人为中心的研究。该数据集规模庞大，包含超过500万帧的标注数据，涵盖7个场景和摄像头，支持常规训练和持续学习两种设置，并包含个人和群体异常行为，如摔倒、踢打等。