JLHI Dataset: A joint localization and human-to-human interaction dataset for indoor environments with different layouts

The dataset is organized into top-level directories named according to participant groups, labeled as Gx, where x represents the group number (e.g., 1, 2, ..., 20). Each group contains data recorded from a unique pair of participants performing human-to-human interaction activities. The group structure enables easy indexing and isolation of samples based on subject pairings and experimental conditions. Within each group folder, the data is further categorized into three subfolders named Layout_1, Layout_2, and Layout_3. Each layout corresponds to a distinct spatial configuration and visibility scenario during data collection. These layouts differ in the number and placement of obstructions between the transmitter and receiver, representing varying levels of occlusion in the environment. Each layout folder contains a single h5 file named CSI_all_samples_Gx_Ly.h5, where x corresponds to the group number and y to the layout number. These files include all samples recorded for that specific group-layout pairing, covering all five activities at three spatial RPs. The dataset folder and file structure are presented in Table 1.
Each H5 file consists of 750 samples, derived from 5 activities at 3 positions with 50 repetitions each. Every sample contains 800 frames collected over 4 seconds at a rate of 200 frames per second, resulting in a total of 600,000 rows per file. The data columns comprise Coordinates, Action, Frame, FrameType, Sample, Amp_dBm[0–63], Phase_deg[0–63], and RSSI_dBm. An overview of each column, along with its corresponding description and example value, is provided in Table 2. Each subject appears in at most two groups and is assigned a consistent role to ensure interaction clarity. In every h5 file, the action column uses integer values from 1 to 5 to represent the five predefined interaction activities. The mapping is as follows: 1 for high five, 2 for handshake, 3 for kick with left leg, 4 for punch with the right fist, and 5 for pushing. This standardized numeric representation ensures uniformity across the dataset and simplifies processing, enabling direct integration with feature extraction pipelines, supervised learning algorithms, and classification models without additional preprocessing.