How we're building the evidence.
DRIVE S(AI)FE is a research project first. This page sets out how data is collected, labelled and tested, and the principles that decide what we will and won't claim.
Six principles govern the design.
They come from the Milestone 4 AI system design, prepared with the Murdoch University project team.
- 01
Vehicle agnostic
The prototype works independently of vehicle hardware, wiring and interfacing systems.
- 02
Edge-first safety response
Current fatigue estimates, immediate alerts and five to ten minute warnings run locally, with no internet needed.
- 03
Post-trip cloud analysis
Detailed recommendations and reports are produced after the trip or on demand. The cloud isn't part of the live warning path.
- 04
Multimodal evidence
Video, physiological, motion, location and operational data are combined. No single measure is treated as definitive.
- 05
Traceable labels
Every training window has a fatigue class, a continuous score and a confidence value, plus data-quality flags.
- 06
Evidence before claims
Results from public datasets are reported separately, and final validation uses drivers and trips the model hasn't seen.
Getting the method right before collecting at scale.
The internal pilot is a test of process. Its job is to find operational problems early, so the field test collects consistent data.
Logistics pilot
Tested the equipment, the driver protocol, data exports and time-alignment across every stream, end to end.
One standard method
Murdoch University is defining a single gold-standard collection method that every participant will follow, so the data is consistent enough to train on.
Real-world data
Drivers from partner fleets record during their normal work, with regular fatigue check-ins, across alert, tired and fatigued states.
No performance claims come from pilot data. Public datasets are used only to benchmark feature extractors and test the pipeline. Model training, calibration and validation will use project test-drive data from Australian heavy-vehicle operations.
What we measure, and why.
Each source has a defined role. The self-rating is a reference for labelling, not something drivers enter while the vehicle is moving.
| Evidence source | Typical measure | Collection point | Purpose |
|---|---|---|---|
| Subjective sleepiness | 1–5 self-rating | Before and after the trip | A general anchor for current sleepiness |
| Driver behaviour | Driver-facing video and independent observer review | Video continuously; selected windows reviewed | Eye closure, blinking, yawning, head pose and visible signs of fatigue |
| Physiological response | Heart rate, pulse intervals and derived HRV | Continuously, where the signal is valid | Change relative to the driver's own baseline |
| Motion and location | Motion sensing and GPS-derived trip variables | Continuously | Trip progress, movement, elapsed time and location context |
| Operational context | Shifts and rosters, breaks, sleep history, time of day, workload | Before, during and after the trip | Prior fatigue risk and context for reading the sensor data |
| Data quality | Automated and manual validity checks | Every analysis window | Missing signals, face occlusion, glare, motion artefacts |
A class, a score and a confidence.
Each valid, time-aligned window gets a discrete fatigue class, a continuous 0–100 score and a confidence value. The self-rating is an important anchor but rarely decides the class on its own. Confidence is lower when a window sits far from a direct rating, or when measures disagree.
Built so no driver appears in both training and testing.
Collect evidence
Driver-facing video, 1–5 ratings, wearable, motion and GPS, and operational context.
Align and label
A common clock and fixed windows, then class, score, confidence and data-quality flags.
Train and validate
Data split at participant level, so results reflect drivers the model hasn't seen.
Deploy
Only features available live are used. The model returns class, score and confidence.
What we'll test before claiming it works.
The final model is chosen on performance with unseen drivers, calibration, false alarms, behaviour with missing data and speed on the edge device, not headline accuracy alone.
Participation is voluntary, and data is handled with care.
- Taking part is voluntary and never a condition of employment. Drivers can pause or stop at any time.
- Participation is governed by a formal ethics framework and informed consent.
- In normal operation the camera feed is analysed on the device and raw video isn't kept.
- Raw video is recorded only where the study protocol, ethics approval and participant consent allow it.
- Only derived trip summaries and fatigue events sync to the cloud.
- Reports are available through role-based access.