Modern vision and learning systems extend the reach of traditional traffic monitoring. We develop and evaluate AI tools that turn dash-cam video, roadside cameras, and probe data into operational signals — vehicle counts, incident alerts, pavement and roadside hazard detections — at a fraction of the cost of fixed instrumentation.

Our emphasis is not on novelty in the model architecture alone but on closing the loop with traffic operations: what false-positive rate is tolerable for an alert that triggers a maintenance dispatch? How do edge-device and bandwidth constraints shape what is actually deployable? How do we evaluate detection performance against the messy ground truth available from agencies?

Scale-CDA is an open-hardware / open-software toolchain that retrofits production cars for generative-AI-assisted cooperative driving automation using under US $1,000 of off-the-shelf parts. An edge-deployed multimodal LLM reads synchronized vision, CAN, and V2X streams and issues structured advisories to verified Openpilot planner hooks — end-to-end decision latency stayed below 60 ms in multi-vehicle road tests, with all inference kept on-board.

Agency platform: DoTPilot

DoTPilot is the lab’s open-source in-vehicle platform for state and county transportation agency fleets. It runs on vehicles an agency already owns and makes each one a two-way link with the agency: an on-board AI dashcam reports roadway conditions as structured findings, while agency feeds — travel advisories, work zone data, incident reports — are delivered back to the driver as in-vehicle warnings. The post-storm damage inspection pipeline, work zone safety warnings, and traffic incident management all run on it.

Selected papers

See the full publications page for the rest.