BasketHAR — Multimodal Basketball HAR Dataset: IMU + Heart Rate + Skin Temperature [arXiv 2025]

Overview

BasketHAR is a novel, domain-specific multimodal dataset for human activity recognition (HAR) in basketball training scenarios, introduced in an arXiv preprint (2025). It was created to address a critical gap in existing HAR benchmarks: the absence of datasets covering complex, professional-level sports actions with simultaneous physiological and kinematic recordings. The dataset features three synchronized data modalities: motion signals from IMUs, physiological metrics (HR + skin temperature), and egocentric video recordings.

Motion data was collected using the MPU9250 sensor embedded in a WT901 SoC worn at the waist front center, providing 3-axis accelerometer, 3-axis gyroscope, 3-axis angle (orientation), and 3-axis magnetic field — all sampled at 200 Hz, the highest IMU sampling rate among public HAR benchmarks. Physiological data was recorded simultaneously from the wrist using an nRF51822 SoC with a Silicon Labs Si1141 sensor, capturing heart rate and skin surface temperature at 1 Hz. Videos were synchronized with sensor data and annotated by basketball-skilled students, with facial blurring for privacy.

The 14 activity classes include both generic actions (Sit, Stand, Walk, Run) and basketball-specific professional movements (Dribble Run, Shoot the Ball, Pass on the Run, Low Dribble with alternating hands, etc.), making it uniquely challenging for conventional HAR classifiers. Baseline results using a multimodal alignment approach (LoRA fine-tuned ImageBind) achieved 78.11% accuracy, outperforming CNN and LSTM models trained on individual modalities.

Column Schema

Key Statistics

• Activity Classes: 14 (generic actions + professional basketball movements)
• Motion Sensors: MPU9250 in WT901 SoC at waist; 200 Hz; 12 channels (acc, gyro, angle, magnetic)
• Physiological Sensors: Si1141 on nRF51822 at wrist; 1 Hz; heart rate + skin temperature
• Video: Synchronized egocentric video (facial blurring applied)
• Annotation: Basketball-skilled students; minimum 2 annotators per label
• Published: arXiv, 2025 — arXiv:2604.17065

Use Cases

• Sports-specific HAR: complex basketball motion classification from waist IMU data
• Skin temperature as a fatigue and thermal load indicator alongside heart rate during training
• Multimodal alignment (video + IMU + physiological) for activity recognition
• LLM-based sports training report generation from translated HAR + physiological data