WSKL/FederatedHAR

Federated Pose HAR是一个面向联邦学习评估的姿态人体活动识别数据集，适用于工业和边缘计算场景。它包含八种不同的上肢姿态，表现为由RGB-D视频记录修改后的FastPose管道提取的13个关键关节点序列。数据集分为集中划分和客户端划分两种形式，方便进行集中训练、本地客户端训练、标准FedAvg和联合集合设置的直接比较。数据集特别适用于隐私敏感环境中的人体活动识别、基于姿态的序列分类和时间序列分类任务。

Federated Pose HAR is a pose-based human activity recognition dataset designed for evaluating federated learning approaches in industrial and edge-computing scenarios. It contains eight distinct upper-body gestures, represented as sequences of 13 skeletal joint keypoints extracted using a modified FastPose pipeline from RGB-D video recordings. The dataset is particularly suited for human activity recognition, pose estimation-based sequence classification, and time-series classification tasks in privacy-sensitive environments.