pretty_name: Tracks
license: cc-by-4.0
tags:
- computer-vision
- human-motion
- robotics
- trajectory
- pose-estimation
- navigation
- retail
task_categories:
- image-feature-extraction
- keypoint-detection
- object-detection
- image-segmentation
- reinforcement-learning
size_categories:
- 1M<n<100M
Tracks Dataset
Real Human Motion for Robotics Planning and Simulation
The ros2 docker container compiliation and visualization script instructions can be found here: https://huggingface.co/datasets/standard-cognition/Tracks/blob/main/keypoint-db/README.md
Overview
The Tracks Dataset captures continuous, real-world human movement in retail environments, providing one of the largest and most structured pose-based trajectory corpora available for robotics and embodied AI research.
Each record represents 3D pose sequences sampled at 10 Hz across normalized store coordinates, enabling research in motion planning, human-aware navigation, and humanoid gait learning derived directly from real behavior :contentReference[oaicite:0]{index=0}.
Key Specifications
| Field | Description |
|---|---|
| Source | Anonymized in-store multi-camera captures (10 retail sites) |
| Scope | ≈ 60,000 hours of human trajectory data (plus 1-hour evaluation subset) |
| Format | CSV schema, ROS 2–compatible via playback plug-in |
| Sampling Frequency | 10 Hz (10 FPS) |
| Pose Structure | 26 keypoints per person per frame (3D coordinates) |
| Environment | Real retail environments with normalized floor layouts |
| Evaluation Subset | One-hour segment including trajectories + store layout |
| Key Metrics | ≈ 2.3 M unique shoppers |
| Anonymization | Face and body suppression; coordinate-only representation |
| Governance | Managed under Standard AI’s data governance policies aligned with GDPR/CCPA and Responsible AI principles |
Integration & Applications
- Distributed in CSV with schema documentation and import notebooks.
- Ready for ROS 2 integration for path planning and human–robot interaction simulation.
- Compatible with Python, PyTorch, and standard reinforcement-learning frameworks.
Example Research Uses
- Motion prediction and trajectory planning
- Reinforcement learning for humanoid gait and control
- Human-aware navigation and avoidance behavior
- Simulation of human–robot interaction environments
Access
The Tracks Dataset is available now for evaluation and licensing.
- Evaluation subset: 1-hour sample under 30-day Evaluation Agreement (private Hugging Face repo).
- Full dataset: 60,000-hour commercial dataset available by request.
For inquiries or licensing: ✉️ [email protected]
Citation
@dataset{standardlabs_tracks_2025,
title = {Tracks Dataset: Real Human Motion for Robotics Planning and Simulation},
author = {Standard Labs},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/standard-labs/tracks}
}