We model feature-matching uncertainty in entropy-regularized optimal transport and propagate it to counting uncertainty under KKT-based analysis. A Transformer-based model predicts OT slack variables, achieving statistical coverage guarantees for count intervals and +38% accuracy improvement.
@article{zhang2026cvic,title={Cvic: Learning Conformal Uncertainty for Video Individual Counting},author={Zhang, R. and others},year={2026},journal={arXiv preprint},}
We develop a decompose-then-aggregate framework for long-horizon video reasoning using MLLMs (GPT-4o), improving coverage under limited context windows. Adaptive visual prompting improves zero-shot video classification and object grounding on the Woven Traffic Safety Dataset.
@article{zhang2025aap,title={When Language and Vision Meet Road Safety: Leveraging Multimodal Large Language Models for Video-Based Traffic Accident Analysis},author={Zhang, R. and Wang, Binhao and Zhang, Jiaxin and Bian, Zhuoran and Feng, Chen and Ozbay, Kaan},journal={Accident Analysis \& Prevention},volume={219},pages={108077},year={2025},publisher={Elsevier},}
We propose a reinforcement learning framework for adaptive scheduling of traffic data collection on battery-powered edge devices, balancing data quality with energy constraints for long-term deployment.
@inproceedings{zhang2023itsc,title={Learning When to See for Long-Term Traffic Data Collection on Power-Constrained Devices},author={Zhang, R. and Han, Wentao and Bian, Zhuoran and Ozbay, Kaan and Feng, Chen},booktitle={2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC)},pages={4676--4682},year={2023},organization={IEEE},}
We introduce pseudo string stability to model long-horizon operation under Bernoulli detector failures. A theoretical link between detection performance and stability in the frequency domain enables cost-effective detection design with provable stability guarantees.
@article{wang2023trc,title={Anomaly Detection and String Stability Analysis in Connected Automated Vehicular Platoons},author={Wang, Yujie and Zhang, R. and Masoud, Neda and Liu, Henry X.},journal={Transportation Research Part C: Emerging Technologies},volume={151},pages={104114},year={2023},publisher={Elsevier},}
We curate 300K+ trajectories from the Lyft L5 Prediction Dataset and implement sequential prediction for vulnerable road users (VRUs), proposing a graph-based Deep Q-Network (DQN) for interaction-aware motion planning at intersections.
@article{zhang2023trc_rl,title={Predictive Trajectory Planning for Autonomous Vehicles at Intersections Using Reinforcement Learning},author={Zhang, Enshuo and Zhang, R. and Masoud, Neda},journal={Transportation Research Part C: Emerging Technologies},volume={149},pages={104063},year={2023},publisher={Elsevier},}
We process 10K+ time-headway profiles from Lyft L5 Prediction Dataset and NGSIM Dataset, applying Bayesian changepoint detection to identify regime shifts in car-following behavior. ANOVA shows a statistically significant reduction in headway variability with AV presence.
@article{zhang2023jtea,title={Impact of Autonomous Vehicles on the Car-Following Behavior of Human Drivers},author={Zhang, R. and Masoud, Saeed and Masoud, Neda},journal={Journal of Transportation Engineering, Part A: Systems},volume={149},number={3},pages={04022152},year={2023},publisher={ASCE},}