HOMI

Human Outdoor Motion Inheritance: LiDAR-Based Agricultural Worker Activity Dataset




Duration
2026 – Ongoing

Researcher
Hanju Seo

Affiliation
- EarthCode Ltd.
- D&AD Future Impact Fund

Role
- Principal Investigator
- Dataset Design (activity scenario definition, sensor placement, field protocol)
- Data Collection (LiDAR field deployment, ROS2)
- ML Pipeline (point cloud processing, feature extraction, classification)

Support
- D&AD Impact Council




Farming is one of the oldest forms of human labour, yet almost none of its bodily knowledge has ever been recorded as data. The posture of squatting to tend a crop, the rhythmic strike of a Ho-Mi against soil, the arc of a watering can across a row: these movements have been passed down through experience alone, absorbed in the body, spoken in fragments, shaped differently by each culture and each generation. They have never been connected to the systems that measure, automate, or scale. HOMI began with a simple question: if agricultural automation and robotic systems are going to work alongside farmers, what data will they need? And who is collecting it?

Existing LiDAR datasets were built for roads and buildings. SemanticKITTI, nuScenes, ScanNet: none of them were designed for an unstructured outdoor environment like a small farming plot. HOMI addresses this gap directly. A single fixed Unitree L2 LiDAR sensor was deployed at a citizen farm in Suwon, South Korea, with no RGB camera. It captured 16,676 point cloud frames across eight activity scenarios: idle standing, walking, static squatting, moving squatting, watering, Ho-Mi tool use, and an unscripted mixed sequence. HOMI is the first ground-level LiDAR dataset of agricultural worker activity in an unstructured farming environment.

The dataset captures something more than motion. Ho-Mi use, a traditional Korean hand plow producing a characteristic rhythmic strike-and-pull at ground level, achieved 92% recall under cross-validation. This is a culturally specific motion that has never appeared in a sensing dataset before. A Random Forest classifier trained on 16 geometric and temporal features reached 95.1% accuracy across all six activity classes. Agricultural ridge structures were additionally detected from the static background scan, with all six ridges successfully identified. The movements that farmers carry in their bodies, passed down without notation, are beginning to be recorded. HOMI is one part of a broader effort to build the sensing infrastructure that the future of farming will depend on.



HOMI field data collection at Tapgong Citizen's Farm, Suwon, South Korea. Single fixed LiDAR sensor, eight activity scenarios, 16,676 point cloud frames.