Tennis Activity Recognition Dataset

Overall, the dataset includes 28,800 labeled windows or instances, providing a balanced benchmark for evaluating machine learning algorithms under cross-subject, cross-session, and cross-device experimental protocols. The dataset supports research in wearable sensing, distributed edge artificial intelligence, real-time human activity recognition, and sports analytics. 2. Dataset content The TAR dataset is organized into a hierarchical file structure to facilitate reproducible evaluation under different experimental protocols, including cross-subject, cross-session, and cross-device settings.

At the top level, the dataset is divided into two folders corresponding to the hardware platforms ("Commercial" and "Custom"). Within each device folder, there are subfolders for each participant (P1–P20), followed by two recording sessions ("S1" and "S2"). Each session contains four synchronized CSV files corresponding to the wearable devices placed on the chest, waist, left wrist, and right wrist.

In addition, the top-level directory contains a "subjects_info.csv" file including the demographic metadata of all participants. • The subjects_info.csv file contains the subject identifier, age, gender, height, weight, dominant hand, and racket sports experience of every participant. • Each sensor_chest.csv, sensor_waist.csv, sensor_wrist_L.csv and sensor_wrist_R.csv file contains the synchronized tri-axial acceleration signals acquired from the corresponding body location.

Every row represents one labeled time window composed of 200 consecutive samples. -------------------------- METHODOLOGY -------------------------- 1. Methodology and data acquisition protocol Before data acquisition, all participants received detailed information about the experimental protocol and voluntarily provided written informed consent for both participation in the study and the public release of the anonymized dataset.

The recording protocol was approved by the Ethics Committee of the Universidad Politécnica de Madrid (UPM) under approval reference ESAFMSI-MGM-DATOS-20260522. During each recording session, participants performed six representative tennis-related activities: Drive, Serve, Backhand one-handed, Backhand two-handed, Lob, and Waiting. Before each activity, the movements were explained to the participants, who were encouraged to execute them naturally according to their own playing technique, speed, intensity, and style.

This acquisition strategy introduces realistic inter-subject variability while maintaining a balanced number of examples for every activity. Each participant completed two recording sessions using both hardware platforms (custom-designed wearable nodes and commercial M5StickC-Plus2 devices). During every session, 60 windows of 200 synchronized samples (4 s at 50 Hz) were recorded for each activity, resulting in 1,440 labeled instances per participant.

After data acquisition, a data curation process synchronized the asynchronous Bluetooth Low Energy data streams received from the four wearable nodes. The synchronized recordings were subsequently segmented into overlapping windows of 200 samples with a 150-sample overlap and assigned their corresponding activity labels, producing a balanced and temporally consistent dataset suitable for benchmarking distributed wearable human activity recognition systems.