Analytica Data Science SolutionsContact
Services
Work
Products
Company
ESC

↑↓ to move↵ to open

← Dr. Aous Abdo

Founder background

Work from Dr. Abdo's time at HP. He did this as an employee of HP, so it is part of his background, not Analytica past performance.

Telling drivers apart from the way they drive

One vehicle, 19,000 kilometers, eleven continuous days, and a rotating crew of drivers. From telemetry alone: how many people drove it, and when did each one take the wheel?

Client
A global automotive manufacturer
Sector
Automotive · Telematics
Role
Dr. Abdo, as an employee of HP
Period
19,000 km over 11 days
Result
78% identification accuracy
driver identification accuracy
78%driver identification accuracy
continuous route
19,000 kmcontinuous route
elapsed
11 dayselapsed

The problem

A single vehicle was driven across continents in eleven continuous days by a rotating crew. The question put to the team was deliberately hard: using only what the car recorded, reconstruct how many distinct drivers there were, and when each of them was driving.

No labels existed. This is an unsupervised problem — the answer has to come out of the structure of the data itself, validated afterward against a driver log withheld during modeling.

What was done

The team fused GPS traces, in-car sensor streams and environmental data into a single feature space, then used principal component analysis to compress dozens of correlated measurements into a handful of parameters that carried the variation.

Clustering in that reduced space separated the driving signatures. Three groups emerged, with some overlap at the boundaries — expected when two people drive similarly on the same stretch of road.

Unsupervised — no labels supplied

Driver signatures separating into three clustersVehicle telemetry begins as a single undifferentiated cloud. Principal component analysis stretches it along the directions carrying most of the variance, and clustering then separates three distinct driving signatures with overlap at their boundaries. Checked afterwards against the true driver log, most observations fall in the correct group; the engagement measured 78 percent identification accuracy. The points drawn here show the method, not the validation data.PC1 · 61% of variancePC2 · 23%Driver ADriver BDriver C
Driver ADriver BDriver C✕misassignedThe model was never told there were three drivers, or which drive was whose. The groups fall out of the structure of the telemetry. Checked afterwards against the true log, most observations land in the right group — the crosses are the ones that do not, and they sit where two people drove the same road the same way. The engagement measured78% identification accuracy; the drawing shows the method, not the validation data.

The result

Validated against the true driver log afterward, the model identified who was driving with 78% accuracy.

Behavioral structure survives in ordinary vehicle telemetry and can be recovered without labels — the same capability that finds unexpected operating regimes in industrial sensor data.

Other applications

The same approach works when the subject is the road instead of the driver: detecting surface degradation, hazardous stretches and changing conditions from ordinary fleet telemetry.

Capability

Unsupervised learning, sensor fusion, signal separation

Discuss a similar problem

If this resembles a problem you're facing, we can walk through how it was built and what it would take in your environment.

Start a conversation