53 — Clusters of Driving Behavior from Observational Smartphone Data

Using accelerometry data from a smartphone, Warren et al are able to determine the direction and acceleration of a vehicle in motion on a road system; and from these datasets, this paper provides a way of clustering driving behaviors based on derivatives such as aggression, speed, and other metrics.

Importantly, accelerometry does not include velocity information ($a$ must be integrated once to get $v$, which is a lossy transformation). To derive speed, GPS data was collected simultaneously.

From the 5-terabyte dataset, the study clustered a cohort of 500 drivers based on a variety of metrics:

First, the study determined the “type” of a road (highway, back road, one-way, etc) from the driving behaviors of the participants. This was used to determine “norms” for driving, deviations from which could be used to establish aggression metrics.

Then, the drivers were clustered by acceleration, deceleration (yes, I appreciate those are the same things), tendency to speed, turning radius, etc. The study found that San Francisco drivers nicely segmented into well-defined clusters based on acceleration and turning speed (rotational acceleration).

By combining these data with information about environment such as traffic, weather, or time of day, I imagine you could get some very interesting metrics about roadway “convenience”: Are there intersections where otherwise conservative, safe drivers have to make quick, aggressive turns? Perhaps these are targets for redesign. Are there places where all types of drivers slam on their brakes? Maybe this is a target for better signage.

And of course, all of this behavioral work is hugely useful when designing the behaviors of self-driving agents. Do drivers need to behave differently in different situations? Certainly. Do you need a distribution of driver types on the road to avoid deadlock? Or is this variance why we get traffic at all?