Project · Active

DriveMotion: a benchmark for driver motion forecasting

PI

What it is

DriveMotion standardizes heterogeneous in-cabin monitoring sources into a single skeleton-motion representation for forecasting driver body movements. The benchmark combines the BATON fleet, a web corpus, and AIDE sources into 400 hours of motion data at 10 Hz with 133 keypoints per frame across 9,010 sequences from 360 drivers — synchronized with vehicle telemetry and camera context — and defines a forecasting task: observe 8 seconds of past motion, predict the next 4.

Why it matters

In-cabin monitoring data comes from incompatible sources — different sensors, different skeleton conventions, different sampling rates — which has made it hard to pool data across studies or compare motion-forecasting models on common ground. DriveMotion’s unified representation and privacy-preserving release turn that fragmented data into a reusable benchmark, with 680,082 forecasting windows available for training and evaluation.

Connections

DriveMotion draws on the same fleet instrumentation as BATON and complements TriDrive’s joint driver-vehicle-road forecasting and DriveDNA’s driving-style work. It sits in the lab’s AI for mobility research thread.

Project team

  • Hao Zhou (PI)
  • Yuhang Wang
  • Chuheng Wei (Purdue University)
  • Jingxin Yang (Stanford University)
  • Xishun Liao (University of Central Florida)

Outputs & releases