AI Research
About
Our AI research focuses on building intelligent, resilient systems for urban and transportation environments. We develop advanced AI models that operate at the edge, coordinate through agentic and multi-agent frameworks, and learn from rich, multi-modal sensor data. By combining generative and decision-support AI with robust sensor fusion, we enable real-time perception, prediction, and autonomous coordination in complex infrastructure systems. We also explore quantum and quantum-inspired AI approaches to accelerate learning, optimization, and large-scale system analysis—pushing the boundaries of what intelligent infrastructure can achieve.
Projects
- Agentic AI
Personalized Federated Learning for Multi-Agent Trajectory Forecasting (UTC x DENSO)
We design personalized federated learning strategies for multi-agent trajectory forecasting across distributed roadside sensors, enabling intersections to collaboratively learn motion patterns without centralizing raw data. Our framework adapts state-of-the-art generative trajectory models into FL and introduces new personalization/aggregation strategies (e.g., metric-aware and representation–head personalization) for non-IID, multi-class traffic behaviors in real deployments. (U.S. Patent Pending)
- Edge AI
FLEET: Real-Time Edge-Deployable Personalized Federated Learning for Multi-Agent Trajectory Forecasting (UTC x DENSO)
A plug-and-play, fully on-edge federated learning framework enabling privacy-preserving, real-time trajectory prediction for heterogeneous road users at urban intersections. The system executes the full learning lifecycle—data collection, training, aggregation, deployment, and inference—directly on edge devices, delivering adaptive, scalable safety intelligence for smart cities. (U.S. Patent Pending)
- Generative AI
M3Track: Language-Driven Multi-Agent Multi-Target Multi-Camera for Referral Tracking (UTC x University of Arkansas)
M3Track is a language-driven, multi-agent framework for multi-target multi-camera tracking in complex urban environments. The system leverages large language model agents to interpret natural language queries, coordinate vision modules, and perform spatio-temporal reasoning across camera networks without manual calibration. By enabling interactive query refinement and open-vocabulary tracking, M3Track advances human-centric, scalable AI systems for smart city safety and monitoring.
- Safety AI
iCityGuardian: Human-Interactive Multi-Camera Tracking for Smart Cities (UTC x University of Arkansas)
A human-interactive, AI-driven multi-camera tracking system designed for city-scale safety and mobility operations. iCityGuardian integrates graph-based trajectory association, vision–language understanding, and conversational AI to enable real-time tracking, forensic search, and resilient monitoring under adverse conditions. By supporting natural language queries and human-in-the-loop reasoning, the system enhances situational awareness, transparency, and operational readiness for smart city safety applications. (U.S. Patent Pending)
Pedestrian Crossing Intent Prediction with Uncertainty-Aware Transformer FusionWe propose a lightweight, socially informed intent prediction architecture that fuses psychologically grounded behavioral streams using a compact Transformer and produces calibrated risk scores via uncertainty estimation. The system is designed for efficient, risk-aware deployment on resource-constrained platforms and supports safer autonomous decision-making in complex urban interactions.
Graph Neural Networks for Traffic Risk Assessment via Surrogate Safety MeasuresWe develop a graph learning framework that converts multi-agent trajectories into spatiotemporal interaction graphs and predicts frame-level traffic risk using interpretable surrogate safety indicators (e.g., TTC, PET, GT). The approach generalizes across intersection and highway environments while supporting low-latency inference suitable for infrastructure-side monitoring.
- Sensor Fusion
CLIFE: Edge-Native Camera–LiDAR Fusion for Reliable Roadside VRU Perception
We develop an edge-native camera–LiDAR fusion pipeline that performs targetless online calibration and lightweight late fusion with integrated multi-object tracking, running fully on embedded hardware. CLIFE was deployed across 12 signalized intersections in Chattanooga and demonstrated robust VRU perception under occlusions, low-light, and adverse weather, while sustaining real-time throughput on edge devices.
AI-Powered Collision Risk Assessment and Post-Crash Detection for Urban Mobility (UTC x DENSO)
We build an edge-computing solution for smart intersections that integrates trajectory prediction and multi-sensor fusion to support collision risk prediction and post-crash detection for vehicles and vulnerable road users. This effort is a joint industry–academia collaboration with DENSO, aiming for scalable real-time deployment and safety impact.
- Quantum AI
Toward Quantum-Enabled Intelligent Transportation Systems via Hybrid Graph Representation Learning
This project explores a hybrid learning framework for traffic forecasting that integrates spatio-temporal graph neural networks with variational quantum circuits. By enriching graph-based traffic representations through quantum-enhanced feature transformation, the approach improves the modeling of complex urban traffic dynamics using both benchmark datasets and real-world sensor data.
A Novel Quantum-AI Hybrid Approach to Large-Scale Traffic SimulationThis project explores the possibility of designing an AI-based traffic simulation platform from the quantum perspective. This study achieves both macroscopic traffic state simulation and microscopic agent-based modeling. Compared to existing simulations, the quantum mechanics in this proposal offers tools to model probabilistic, non-deterministic, and highly entangled systems, which can mirror real-world vehicle behavior under complex traffic conditions.