Case Study · Edge AI

Driver monitoring. Fully on-device.

A real-time embedded vision system that evaluates fatigue, attention and posture directly on the target hardware and turns the results into clear, graded driver alerts.

AutomotiveMediaTek Genio 700 · GRINNDEEPX AI acceleration

Driver monitoring dashboard with live camera analysis and driver status

The challenge

Driver monitoring must interpret several weak signals at once and react quickly without depending on a cloud connection. The application needed to combine camera capture, face and body analysis, calibration, configurable alert logic and an operator-ready interface on one embedded platform.

Engineering approach

Crossware implemented the solution on a MediaTek Genio 700 platform with a GRINN board, a DEEPX AI accelerator and a 1920 × 1080 USB camera. The processing pipeline tracks facial landmarks, eye state, head pose and body keypoints. Temporal logic converts these signals into stable fatigue and attention states while configurable persistence and recovery periods prevent alerts from flickering.

40–41 fpsFace analysis rate on the target

28–35 msCamera frame to face result

68 + 17Facial landmarks and body keypoints

0 cloudAll inference and alert logic run locally

System demonstration

See the Driver Monitoring System in action.

The demo shows the live on-device analysis, driver-state evaluation and visual feedback running together on the embedded target.

Application views

From live inference to explainable alerts.

The interface exposes the underlying signals, system state and event history so developers can validate behaviour and adapt thresholds to the vehicle context.

Simplified driver monitoring view — **Focused driver view.** Live landmarks, attention status and the essential driver-state indicators.

Critical microsleep alert in the driver monitoring application — **Microsleep detection.** Eye closure over time raises a clear critical event.

Head drop warning in the driver monitoring application — **Head-drop detection.** Pose and eye-state information distinguish a momentary movement from a safety-relevant condition.

Configurable alert thresholds for driver monitoring — **Configurable behaviour.** Thresholds and timing can be tuned for the intended vehicle and validation programme.

Detection scope

Multiple signals, one consistent driver state.

Critical

Immediate safety events

Microsleep, prolonged eyes-off-road and an unresponsive head drop.

Warning

Degrading attention

Looking away, head drop, high fatigue and low attention scores.

Caution

Early indicators

Yawning, leaning, hand-off-wheel, hidden face and initial signs of fatigue.

The alert logic combines current observations with recent behaviour. Persistence and recovery windows suppress short-lived noise, while combined fatigue and posture signals can increase the alert level.

Validation and integration

Built as an engineering tool and a product foundation.

Driver monitoring source, display and overlay configuration — **Source and overlay control.** Camera, display mode and diagnostic overlays are configurable in the application.

Driver monitoring replay mode during calibration — **Replay and calibration.** Recorded sequences support repeatable development and validation without requiring every scenario in the vehicle.

Driver monitoring system ignoring a passenger face — **Driver selection.** The system locks onto the driver and ignores additional faces in the cabin.

Detailed driver monitoring engineering dashboard — **Engineering dashboard.** Performance, inference and driver-state data remain visible during system tuning.

Planning an embedded vision product?

We can help with camera pipelines, accelerator integration, inference optimisation, application architecture and target validation.

Learn more →