Publications
Three new papers: PMA-Diffusion, Weaver, and TrajFlow
I’m excited to share three new papers from the lab, all on generative and deep-learning methods for traffic modeling.
PMA-Diffusion: A Physics-guided Mask-aware Diffusion Framework for Traffic State Estimation from Sparse Observations has been published in Transportation Research Part C: Emerging Technologies (Volume 190, Article 105801). Led by Lindong Liu, together with Zhixiong Jin, this work reformulates sparse traffic state estimation as a Bayesian inverse problem. Instead of requiring fully observed training data, PMA-Diffusion learns a mask-aware diffusion prior directly from incomplete speed fields and applies weak traffic-physics guidance only at inference through an adaptive anisotropic smoothing projector. On the I-24 MOTION dataset it reconstructs complete highway speed fields even under severe sparsity—outperforming representative baselines at as little as 5% visibility.
Weaver: Kronecker Product Approximations of Spatiotemporal Attention for Traffic Network Forecasting, by Christopher Cheong, Gary Davis, and me, has been accepted and is in press at Transportation Research Part C: Emerging Technologies. Weaver uses the mixed Kronecker matrix–vector identity to approximate spatiotemporal attention efficiently, factorizing spatial and temporal structure to scale attention-based forecasting to large traffic networks.
TrajFlow: A Generative Framework for Occupancy Density Estimation Using Normalizing Flows, by Mitch Kosieradzki and me, has been accepted and is in press at Artificial Intelligence for Transportation. TrajFlow casts occupancy density estimation as a generative modeling problem, using normalizing flows to learn flexible, tractable densities over future trajectory occupancy.
Congratulations to Lindong, Chris, Mitch, and all of our collaborators on these milestones! 🎉