PedRisk estimates the chance of a pedestrian–vehicle conflict from tracked movement and raises a low, medium or high alert ahead of the closest approach. Built on 41 real runs from the FHWA Intersection Safety Challenge.
Choose a run to play back the tracked vehicle and pedestrians with the model's risk score at each moment.
Each run is scored by a model trained without it. Lanes and crosswalks are fitted to all 41 vehicle and pedestrian tracks and styled after the site's aerial view. It is a schematic, not a surveyed map; signal heads are not shown.
Adjust the eleven model inputs, or start from a recorded run.
FHWA Intersection Safety Challenge, Stage 1B validation data. Each run has a ground-truth track every 0.1 s for a vehicle and one to three pedestrians or riders, plus a conflict / no-conflict label.
Runs containing each type (no conflict / conflict).
| Feature | Meaning |
|---|
Standardized coefficients. Right of center raises the conflict probability.
80% train / 10% validation / 10% test by run, rotated over all 41 runs and repeated with five shuffles (mean ± sd). Only 12 runs are conflicts, so read these as a proof of concept.
| Model | Precision | Recall | F1 | Accuracy | AUC |
|---|
Distance rule: alert when the closest separation is under a threshold learned from the training runs.
Full sensor data: FHWA on data.gov.