Research

With the growing deployment of vehicular communication technologies (e.g., V2X and 5G cellular), vehicles and infrastructure are becoming connected to each other and to the cloud. This connectivity enables a fundamental shift: rather than confining 3D data operations to individual vehicles or sensors, 3D data can now be streamed, aggregated, interpreted, and delivered across the network. This shift introduces new opportunities but also new challenges: not just how to sense or store 3D data, but how to operate and manipulate it effectively in a cloud-connected vehicular ecosystem.

My research explores how 3D data can be harnessed as a shared and networked resource in vehicular systems:

• [Reconstruction] How to collect and reconstruct 3D traffic scenes by aggregating data from distributed sensors
• [Interpretation] How to extract semantic understanding from the reconstructed scene in the cloud in real-time
• [Delivery] How to deliver rich 3D content from the cloud to vehicles to support immersive in-vehicle visualization applications

Ultimately, my research envisions a future where 3D data is no longer treated merely as a local sensing artifact, but as a shared medium that enables collaborative perception, coordinated decision making, and immersive vehicular experiences.

Reconstruction

RECAP: 3D Traffic Reconstruction

RECAP focuses on vehicle-to-cloud 3D sensor data streaming and explores how vehicles equipped with onboard 3D sensors can offload their data to the cloud while the cloud enables high-fidelity reconstruction of complex traffic scenes via multi-view fusion from inputs from multiple moving vehicles.
The core challenge in RECAP lies in accurately and quickly fusing large volumes of 3D data from dynamic, sparsely overlapping views across space and time. Building upon prior work on 3D view fusion, RECAP introduces techniques that minimize reconstruction error and computational cost in these highly asynchronous observation scenarios.

Interpretation

Cooperative Infrastructure Perception

CIP targets infrastructure-side sensing where 3D sensor-equipped roadside units stream 3D data to the cloud for large-scale traffic perception. It investigates how to extract semantic insights from traffic intersections in real-time.
To generate perception outputs fast enough for autonomous driving use cases, CIP must process large scale 3D scenes within 100 ms while achieving accuracy that match or exceed state-of-the-art vision algorithms. To meet this requirement, CIP introduces a novel alignment algorithm for accurate view fusion, along with efficient implementations of key perception tasks including dynamic object extraction, tracking, and motion estimation.