Research.

Multi-camera vision for gait analysis and fall-risk assessment

How can visual representations of human movement remain useful across people, camera systems, and everyday environments?

My current research at Brown LEMS connects the detection of acute falls with the broader goal of understanding changes in mobility and balance as people age. Fall detection provides a focused starting point for studying movement over time, with gait analysis and fall-risk assessment guiding the broader research direction.

I treat data and evaluation design as central research questions. Staged recordings support controlled experiments, but studying natural movement in older adults also requires data that reflect the intended population and setting. I address these complementary needs through B-Gait, an extensible experimental testbed, and BROWN-BEAR, a real-world dataset initiative.

I independently lead this project within a cross-lab initiative on AI-driven healthcare. My work spans research formulation, benchmark design and implementation, experimental evaluation, and the design and validation of a data-acquisition prototype. I am advised by Prof. Benjamin B. Kimia, with supervision from Chiang-Heng Chien.

B-Gait

B-Gait grew out of a question about experimental comparability: what does a reported result mean when data preparation, movement representations, and evaluation rules differ? I designed and built a testbed that makes these choices explicit. Datasets, camera views, representation pipelines, decision models, and evaluation protocols are declared within a common experimental framework.

My goal is to make B-Gait a reusable research platform for fall detection. Its extensible design supports introducing new datasets and methods while keeping experimental assumptions visible, so researchers can define, reproduce, and build on comparisons within a shared framework.

BROWN-BEAR

I designed BROWN-BEAR (Brown Real-World Observation for Wellness and Neuromotor Behavioral Evaluation and Activity Recordings) dataset to address the gap between staged activity recordings and the natural movement of older adults. The dataset initiative aims to capture unscripted, multi-view observations for studying gait, falls, and mobility changes, grounding future method development and evaluation in the population and conditions that motivate the research.

I developed the acquisition design and validated the prototype. BROWN-BEAR is now at the prototype stage, awaiting real-world deployment.

Longer-term direction

I aim to build on this experimental and data foundation to develop reliable pose estimation and kinematic analysis, connecting observable movement with changes in mobility and balance and, longer term, the assessment of neurodegenerative disease.

Time-series anomalies and risk modeling

How can models learn from rare events and characterize the patterns behind them?

I contribute to collaborative research on credit risk, class imbalance, and time-series anomaly profiling. My contributions combine involvement in method design with code implementation, experiments, and analysis, connecting modeling ideas with empirical evaluation.

Our credit-risk work models repayment behavior over time, using attention and residual connections to retain information from both recent and distant observations. Our oversampling study examines how dataset characteristics and classifier choice affect the value of resampling. Together, these studies investigate how temporal context and data composition shape predictive performance.

My current collaborative work extends anomaly detection toward profiling observable patterns and underlying mechanisms. CDAP combines normal-reference calibration with intervention-guided concept disentanglement to reduce spurious responses to benign fluctuations and separate responses to different anomaly concepts when type annotations are scarce. The manuscript is under second-round review at AAAI 2027.

Low-dose CT denoising

How can diffusion-based restoration preserve anatomical detail at a practical computational cost?

I study efficient medical image restoration through two connected design questions: how to make diffusion-based denoising less computationally demanding, and how to suppress noise while retaining fine anatomical structure. In MAN, I developed a compact latent restoration pipeline that combines a perceptually optimized autoencoder with conditional diffusion and efficient deterministic sampling.

TAFG-MAN builds on this foundation by examining when different image cues should guide restoration. Its timestep-adaptive frequency gates emphasize stable structural information early in denoising and progressively introduce high-frequency detail at later stages. This design improves detail preservation and perceptual quality over the base model at essentially the same inference cost.

I independently developed this research from conception and method design through implementation, experiments, and manuscript writing, advised by Prof. Xiangjian He with supervision from Jingxi Hu. This work brings representation design, computational efficiency, and detail-aware conditioning into a unified approach to medical image restoration.