Dataset¶
HLS-HAR was designed for long-session sports recording rather than isolated window classification. Huawei sports watches captured wrist acceleration and angular velocity in real sports-health monitoring sessions. Research access is coordinated by the Hainan University organizers.
Request research access
HLS-HAR participant recordings are not distributed on GitHub. Before the planned PhysioNet release, researchers can submit the dataset request form. The Hainan University data organizer reviews each request.
Activities¶
The foreground vocabulary contains badminton, rope skipping, dumbbell fly, running, and table tennis. Background motion is modeled internally but is not reported as a workout record.
| Activity | External-test segments |
|---|---|
| Badminton | 32 |
| Rope skipping | 20 |
| Dumbbell fly | 20 |
| Running | 20 |
| Table tennis | 22 |
| Total | 114 |
Splits¶
The split is fixed at the recording level.
| Split | Recordings | Use |
|---|---|---|
| Training | 80 | model fitting and training-stage selection |
| Development / calibration | 20 | temporal-policy tuning and diagnostics |
| Independent external test | 37 | final scoring after the operating point is frozen |
The 100-record model-development pool contains 284 labeled activity segments, about 34.8 million samples, and 130.5 hours of sensing. The 80-record training split contributes about 28.2 million samples and 113.5 hours.
Development-set scores are diagnostics, not independent estimates. External labels are used only for final scoring; they are not used to select checkpoints, fusion rules, TRL parameters, or reported variants.
Signals¶
The reported system uses six 100 Hz channels:
| Signal | Meaning |
|---|---|
| \(a_x, a_y, a_z\) | tri-axial wrist acceleration |
| \(\omega_x, \omega_y, \omega_z\) | tri-axial wrist angular velocity |
| timestamp | millisecond reference for segment boundaries |
Three overlapping views are constructed with a one-second step:
| Window | Samples | Training windows |
|---|---|---|
| 3 s | 300 | 281,871 |
| 5 s | 500 | 281,711 |
| 8 s | 800 | 281,471 |
Labels¶
Each annotation is an (activity, start, end) segment. Evaluation therefore
measures complete records, including event count, boundaries, and duration.
Predictions are matched one-to-one with same-class labels at IoU > 0.5.
Scope¶
Public HAR datasets only partly overlap with this setting. Some use phone accelerometers, short scripted windows, daily activities, or no gyroscope; others do not provide long-session segment records. Cross-paper window accuracy is therefore not directly comparable with HLS-HAR segment F1.
HAR70+, WISDM-phone, PAMAP2, and OPPORTUNITY are used only for TRL portability checks with dataset-specific models and parameters. They are not leaderboard comparisons and do not establish transfer of the HLS-HAR result.
Access and privacy
Participant recordings are not stored in the public repository. Before the planned PhysioNet release, research requests follow the dataset access instructions. Storage layout and asset checks are documented in Assets.