AFUN Human Activity Recognition Dataset for Balance Based Static Pose Classification using ARKit

The dataset contains joint-coordinate samples generated using ARKit body tracking. Data collection was undertaken on an iPad Pro 11-inch (4th generation) running iOS 18.0. Each sample contains the x, y, and z coordinates of 91 ARKit-tracked body joints, producing 273 numerical input features, plus one categorical pose label. The dataset does not contain image pixels, still images, video frames, raw camera footage, or child gameplay data. The dataset contains 2,421 original samples collected from 12 adult participants. Participants performed 11 labelled pose classes across six camera-relative locations, with three repetitions per pose/location combination. The dataset is provided in two organisational formats. The file AFUN_HAR_Dataset.csv contains the full compiled dataset, while the participant-level files, such as P1_AFUN.csv, P2_AFUN.csv, etc., contain the same data separated by participant. These participant-level files are included to support reproducible participant-level splitting and leakage-aware model development workflows. The dataset was designed as an ARKit-native feasibility dataset rather than as a general-purpose HAR benchmark. Its purpose is to support research on lightweight, on-device pose classification using the same skeletal-coordinate representation available within the AFUN Platform during deployment. It is intended to support the reproducibility of the thesis model development pipeline and future research on ARKit-compatible adaptive AR interventions.