Problem
A ride-hailing and delivery platform wants to know which of its drivers drive unsafely, before an accident happens. Consumer complaints cover only transport rides and still miss most drivers, and delivery verticals get none. Fitting every vehicle with a dedicated inertial measurement unit (IMU) costs too much at this scale. The only sensors every driver already has are in the phone, which sits at an unknown angle and gives noisy readings, especially on motorcycles.
Example from the article: Grab turns phone sensor data from driver and delivery partners into daily counts of harsh braking, acceleration, cornering and speeding, shown in a personal Driving Safety Report in the driver app.
Solution in brief
Rotate raw phone accelerations into the vehicle frame, find peaks above thresholds, and drop events that fail sanity checks. Speeding comes from GPS pings compared with road speed limits, or with crowd-derived speed norms where limits are unknown. Aggregated counts feed a report, a safe driving score and a SHAP-based hint about what to fix.
Questions this case answers
- How to detect unsafe behaviour of mobile workers at scale from noisy smartphone sensors when no dedicated hardware and few direct feedback signals are available
- How to derive per-segment speed norms from crowd GPS pings when the reference limits are missing or incomplete
- How to detect harsh braking, acceleration, cornering and speeding from drivers' phones without installing hardware in vehicles?
- How to recover vehicle acceleration from a phone that lies at an arbitrary angle in the car?
Tools and methods
- Euler-angle rotation with PCA — recover vehicle-frame acceleration from phone accelerometer data
- Low-pass filter, variational decomposition, Kalman filter — denoise by noise level
- Peak finding with thresholds — detect harsh acceleration, braking and cornering
- Speed profiles from GPS pings — set speeding thresholds per road and hour
- SHAP — find the manoeuvre that most raises the risk of feedback or accident
Google or a one-time email code. No password.