EnjoyPoomsae: A Pose-Based Taekwondo Poomsae Forms Dataset

This dataset, EnjoyPoomsae, contains pose-based skeletal time-series data extracted from Taekwondo Poomsae form executions (Forms 1–4). The source videos were collected from publicly available online resources and include unconstrained real-world recording conditions such as camera motion, multiple persons in the scene, and varying viewpoints. To ensure consistent subject selection, a multi-stage processing pipeline was applied including human detection, multi-object tracking, and pose estimation. Only one tracked subject per clip was retained. For each video clip, 2D body keypoints were extracted and converted into temporal sequences. To improve invariance and model robustness, two geometric normalization strategies were applied: Translation normalization: the nose joint is shifted to the origin (0,0) Scale normalization: the torso length (shoulder-to-hip distance) is scaled to one Four dataset variants are provided: no normalization translation normalization only scale normalization only combined translation and scale normalization To reduce sequence length and memory requirements while preserving motion structure, temporal frame subsampling with stride=4 was applied during preprocessing. Each sample is stored as a NumPy NPZ file containing: X: pose keypoint sequence array (T × D) y: class label (form ID) The dataset contains 540 samples with balanced class distribution. Each class is performed by different individuals to improve subject independence. The full preprocessing, feature extraction, and training pipelines are publicly available in the associated GitHub repository (see Related Works section).本数据集名为EnjoyPoomsae，包含从跆拳道品势（Poomsae）第1至4套演练中提取的基于姿态的骨骼时序数据。 其源视频采集自公开在线资源，录制环境为无约束真实场景，涵盖相机运动、多人物出镜、视角多变等情况。为确保一致的受试对象选取，本研究采用多阶段处理流水线，涵盖人体检测、多目标跟踪与姿态估计环节，最终仅保留每个视频剪辑中追踪到的单一受试对象。 针对每个视频剪辑，本研究提取2D人体关键点并转换为时序序列。为提升模型的不变性与鲁棒性，采用两种几何归一化策略： 1. 平移归一化：将鼻尖关节平移至原点(0,0) 2. 尺度归一化：将躯干长度（肩髋间距）缩放至单位长度 本数据集提供四种变体： - 无归一化 - 仅平移归一化 - 仅尺度归一化 - 平移与尺度联合归一化 为在保留运动结构的前提下缩短序列长度、降低内存占用，预处理阶段采用步长为4的时序帧下采样操作。 每个样本以NumPy的NPZ文件格式存储，包含以下数据： - X：姿态关键点序列数组，维度为T × D - y：类别标签，即品势套数ID 本数据集共包含540个样本，各类别样本分布均衡。每个类别均由不同受试对象完成演练，以提升受试对象独立性。 完整的预处理、特征提取与训练流程已在关联GitHub仓库公开，详见相关研究部分。