Research
I carried out my doctoral work in the Feng Group at the University of Michigan-Dearborn, where I completed my Ph.D. in Industrial and Systems Engineering in August 2026. Each project builds a tool for road safety research and tests it against a reference outside the model (an ultrasonic proximity sensor and Waymo LiDAR for passing distance, hand-labelled events for overtaking detection, held-out sites for the vehicle classifier, and the EPA test cycles that vehicles are certified against).
Since the Ph.D.
Does Generated Driving Video Keep Its Camera?
Generated driving video is now used as simulation data, both to train perception models and to evaluate planners. I built a ground-plane check that tests whether a generated clip keeps the camera it was conditioned on: a detector only finds cars, and the camera height implied by each car's box and the horizon is arithmetic, with no learned model inside the test. I applied it to the released clips of six generators on 150 nuScenes validation scenes, paired by scene and by car with the real footage.
Key results
- The check flagged one generator, DiST-4D, whose clips imply a camera height 19.1% below the real clips. An image alignment points to a fixed vertical crop and stretch of the frames, and undoing it brings the difference to -0.25%.
- The other five generators sit between 0.4% and 5.0% below the real clips. On real clips the check detects an injected 1° vertical shift and moves by 0.2% or less under a ±10% zoom.
- A difference in implied height can come from a camera change, a crop, or cars drawn at a different size, and the check does not separate these.
Accepted, UrbanState: Multimodal Urban Sensing and Representation for Autonomous Mobility (Workshop at WACV 2027)
Doctoral Research
Vision-Based Lateral Passing Distance Estimation from Bicycle-Mounted Cameras: A Projective Geometry Approach
A framework that estimates how closely a vehicle passes a cyclist from 2D bounding-box geometry. Three features derived from the pinhole camera model are fitted against an ultrasonic proximity sensor, which serves as the training signal.
Key results
- 0.126 m mean absolute error (MAPE 8.2%) on 92 passing events, leave-one-group-out cross-validation.
- Re-fitted on the Waymo Open Dataset with LiDAR distances as supervision (a second platform and a longer range): R² = 0.812, MAPE 6.6%, across 11,248 detections.
Poster accepted, International Cycling Safety Conference 2026; journal manuscript in preparation
A Geometry-Informed Computer Vision Method for Detecting and Examining Overtaking Vehicles From a Bicycle
An automated pipeline that detects vehicle overtaking events from a single rear-facing bicycle-mounted camera, with no multi-sensor rig and no calibration. RT-DETR detects vehicles and a multi-object tracker follows each vehicle across frames. Three checks from camera geometry (bearing-angle trend, apparent-size growth, spatial confirmation) then keep only the tracks whose movement matches a vehicle passing from behind.
Key results
- 98.1% recall across 315 annotated events.
- Mean advance warning of 2.37 s before passage; 83.2% of events exceed a 1.5 s reaction-time threshold.
Quantifying Drivers-Overtaking-Bicyclists with Surrogate Safety Measures Derived from High-Resolution Digital Lidar
A lab study of how drivers overtake bicyclists, using surrogate safety measures computed from high-resolution LiDAR.
TRB Annual Meeting 2026 (R. Moustafa, G. Padmanaban, F. Feng) · NSF PAR
A Multi-View Vehicle Image Dataset and Two-Stage Pipeline for Fine-Grained Vehicle-Type Recognition at Ground Level
A two-stage pipeline (RT-DETR detection plus a fine-tuned ViT-Base/16) that classifies passing vehicles into six categories defined by injury risk to cyclists (passenger car, SUV, pickup, minivan, large van, commercial truck), rather than make and model. I assembled and labeled the custom dataset and designed the taxonomy. The classifier outputs "unknown" when its confidence is low. I have started testing whether conformal prediction can replace the abstention threshold I chose empirically.
Key results
- 94% accuracy in-distribution (3,805 events; per-class F1 0.91-0.97).
- 89% accuracy on separate held-out sites with no retraining, with abstention correctly rising under distribution shift.
In preparation · Earlier preprint · Code
A Machine Learning Framework to Identify Aggressive Driving Based on Vehicle Kinematics and Driver Pedal Operations
A machine learning framework that classifies aggressive versus non-aggressive driving from longitudinal jerk and driver pedal operations. The core idea is a speed-adjusted jerk threshold, four speed segments aligned with FHWA roadway functional classes, jointly optimized with the classifier over a large grid search, because the same jerk magnitude means different things at 20 mph and 60 mph.
Key results
- 94% accuracy, AUC-ROC 0.971 (Random Forest), across 556 trips and 7 vehicle models.
- Speed-adjusted thresholds consistently beat fixed thresholds across all classifiers and feature sets.
Abstract accepted, SAE WCX 2027
A Comparative Analysis of Acceleration and Deceleration Profiles for Aggressive Driving Styles and Fuel Economy Test Cycles
The same reliability question, asked of a test procedure instead of a model: does the standardized cycle used to certify vehicles represent the driving it certifies? A speed-segmented characterization of real-world aggressive and non-aggressive driving, benchmarked against four EPA fuel-economy test cycles. The US06 cycle approximates aggressive acceleration but substantially overstates aggressive deceleration intensity; all milder cycles also overestimate deceleration relative to on-road driving, with implications for powertrain calibration and fuel-economy labeling.
SAE Technical Paper 2025-01-8605
WCX SAE World Congress Experience, 2025
Earlier Work
Computational Human Performance Modeling using Queuing Network in an Open-Source Platform
An open-source implementation of the QN-MHP (Queuing Network-Model Human Processor) cognitive architecture in Python and SimPy, modeling human multitask performance as a queuing network of information processors so that interface designs can be evaluated computationally. Validated against empirical visual-manual task data.
An Autonomous Driving System: Dedicated Vehicle for People with ASD and their Caregivers
A human-centered design study for an autonomous vehicle tailored to individuals with Autism Spectrum Disorder and their caregivers, developed through stakeholder engagement, with a customized interior concept and companion application addressing safety and monitoring.