AI & ML interests
AI for transportation
Recent Activity
Papers
CANTraj: Recovering Vehicle Trajectories from CAN Signals in Naturalistic Driving
DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification
Mobility Optimized by Traditional Ideas and Frontier technologies
Website · Datasets · Publications · GitHub · YouTube · Join us
We are a transportation engineering research lab in the Department of Civil and Environmental Engineering at the University of South Florida (Tampa, FL), led by Dr. Hao Zhou.
We improve mobility by combining traditional ideas, such as traffic flow theory and safety modeling, with frontier technologies, such as emerging driving technologies and AI. We believe technology works best when domain knowledge guides it.
What we work on
- Traffic flow theory and operation: car following, string stability, capacity, and how new kinds of vehicles change traffic.
- Emerging vehicle technologies: how production ADAS (adaptive cruise control, lane keeping assist), automated vehicles and EVs actually behave on public roads.
- AI for mobility: vision-language models, world models and AI agents for driving safety and traffic operations.
Datasets and benchmarks
Every dataset and model on this page was led by Yuhang Wang, a PhD student in our lab and the first author of each paper. Each release was recorded in production vehicles, driven by real people on public roads.
| Release | What it captures | Scale | Venue | Access |
|---|---|---|---|---|
| ADAS-TO | Human takeovers during ADAS use: the moment control comes back to the driver | 16,446 clips · 364 drivers · 179 vehicle models | ITSC 2026 | Gated · CC BY-NC 4.0 |
| BATON | Control transitions in both directions (handover and takeover), with road and cabin video, CAN and radar | 3,593 transitions · 162 h · 150 drivers | KDD 2027 | Gated · CC BY-NC 4.0 |
| DriveDNA | Driving style identification, designed so models can't cheat by recognizing the car or the road | 975 h · 4,121 drives · 465 drivers | KDD 2027 | Gated · research license |
| DriveMotion | Whole-body driver motion (133 keypoints at 10 Hz) for forecasting; skeletons only, no cabin video | 400 h · 9,010 sequences · 360 drivers | WACV 2027 | Gated · CC BY 4.0 |
| VLAlert-Bench | When a driving system should stay silent, keep watching, or alert the driver | 13,534 videos · 192,892 labeled 1-s ticks | CoRL 2026 | Gated · CC BY 4.0 |
| OpenLKA | How production lane keeping assist performs in real traffic | 389 h · 62 vehicle models · 53 drivers | ITSC 2025 | On request |
Smaller subsets: ADAS-TO-Sample · ADAS-TO-Critical · BATON-Sample · DriveDNA-Sample
How to get access: most releases are gated. Click "Request access" on the dataset page and we review research requests by hand. BATON-Sample and DriveDNA-Sample are approved automatically, so they are the fastest way to try the data.
Models
- TriDrive: forecasts the driver, the vehicle and the road together, and runs in the car at 5 Hz for driver monitoring. Code and checkpoints, gated. (paper)
Join us
We are recruiting motivated PhD and MS students. See motiflab.net/join.