Code & Data
Start with the Bayesian calibration guide, stochastic driving behavior, or traffic simulation guide.
Implementations behind my papers, plus an interactive simulator. Everything here is public on GitHub; issues and pull requests are welcome.
Research code
When Context Is Not Enough
Couples a context-dependent car-following mean with temporally correlated Gaussian-process residuals. Includes CSV validation, model training, stochastic simulation, and a synthetic quickstart for use with your own leader-follower trajectories.
The source code is released under the MIT License. No highD data, pretrained weights, or paper benchmark results are bundled.
IDM_Bayesian_Calibration
Implements MA-IDM and Bayesian IDM, the dynamic AR+IDM formulation from the Bayesian dynamic regression paper, and multi-vehicle ring-road simulations. Models are built with PyMC and calibrated on the highD naturalistic driving dataset; preprocessed trajectories ship with the repository.
Gaussian_Velocity_Field
Reference implementation of the Gaussian velocity field (GVF), which represents multi-vehicle interaction inside a region of interest around the ego vehicle. Each frame becomes a 13 × 17 × 2 tensor over a meshed ROI, giving a fixed-size description of a scene that varies in the number of surrounding vehicles.
Multivehicle-Interaction
Collects results and media for the general framework of learning multi-vehicle interaction patterns from video, and for the Foundations and Trends review on social interactions in autonomous driving.
Interactive demo
traffic-sim
Cars follow the Intelligent Driver Model on a circular track, and you can switch between three driver-noise models drawn from recent work on Bayesian car-following calibration. Small perturbations grow into stop-and-go waves with no bottleneck present, reproducing the Sugiyama experiment. Inspired by Dr. Martin Treiber's work on stochastic car-following.
Datasets remain subject to their original owners' access requirements and terms. Access to the highD dataset must be requested from its provider. Data contents differ by repository; check each README for details. The When Context Is Not Enough release includes no raw or preprocessed highD data.