연구 분야

Research

Intelligent Signal Processing

Multi-channel 1D Biomedical Signal Processing is a subfield of digital signal processing that jointly analyzes one-dimensional time-series bio-signals simultaneously acquired from spatially distributed sensors using mathematical and statistical models. While single-channel processing primarily focuses on temporal variations in frequency and amplitude, multi-channel processing incorporates spatial correlation to model underlying spatio-temporal dynamics. Representative bio-signals in this domain include High-Density surface electromyography (HD-sEMG), electroencephalography (EEG), electrocardiography (ECG), and multi-site photoplethysmography (PPG/rPPG).

We develop non-invasive sensing pipelines to capture cardiovascular dynamics and neuromuscular intent with high fidelity. For vital monitoring, we utilize camera-based remote photoplethysmography (rPPG) to extract facial blood volume pulses, robustly estimating heart rate, HRV, and respiration while actively suppressing motion and lighting artifacts.
To decode motor intent, we process surface electromyography (sEMG) signals by measuring muscle action potentials across electrode arrays. By filtering volume conduction noise and cross-talk, our framework resolves fine motor commands and joint torque in real time, serving as an intuitive, low-latency control stream for next-generation XR interfaces and assistive prosthetics.

We develop intelligent clinical decision-support systems focused on objective diagnosis and quantitative grading in wound care—specifically targeting pressure injuries (bedsores), burns, and diabetic foot ulcers (DFU). Grounded in our proprietary, large-scale standardized clinical datasets, our diagnostic frameworks integrate multi-spectral feature extraction with morphological tissue segmentation to evaluate wound progression, tissue viability, and healing trajectories.
To overcome critical clinical bottlenecks—such as rare pathological cases, class imbalances, and privacy-constrained medical data collection—we research advanced Generative AI architectures (e.g., conditional diffusion models and generative adversarial frameworks). By synthesizing physiologically consistent, high-fidelity synthetic wound lesions and counterfactual progression patterns, we establish robust data augmentation pipelines that significantly enhance model generalizability and diagnostic fairness in real-world clinical environments.

3D Multi-Scale Pose Estimation & Clinical Motion AnalyticsWe specialize in high-precision 3D pose estimation by fusing stereo vision and RGB-D depth sensor pipelines, reconstructing physically accurate metric coordinates (X, Y, Z) across multiple anatomical scales: full-body kinematics, fine-grained hand articulations, and dense facial landmarks. By effectively mitigating monocular depth ambiguity and self-occlusion, our framework translates spatial biomechanics into quantifiable clinical indicators, including spatiotemporal gait abnormality analysis for dementia screening and millimeter-level symmetry evaluation for facial palsy grading.

We investigate Edge-Cloud Collaborative Intelligence architectures that optimize LLM inference by partitioning autoregressive decoding between client browsers and cloud servers. Leveraging client-side WebGPU acceleration, the system executes initial transformer layers locally with zero installation, preserving user privacy and local KV caches while offloading computation-heavy layers to the cloud.
To minimize token generation latency, we employ activation compression and adaptive layer-partitioning strategies. By quantizing intermediate boundary states, the framework substantially reduces per-token network payloads without sacrificing perplexity, ensuring scalable, low-latency token generation across heterogeneous client devices.

We develop model compression and hardware-aware optimization frameworks to deploy deep learning onto resource-constrained embedded systems. By applying Post-Training Quantization (PTQ), Quantization-Aware Training (QAT), and structured pruning, we convert complex networks into low-bit formats (INT8/INT4) with minimal accuracy degradation.
To maximize throughput and energy efficiency, we co-design these models for target Neural Processing Units (NPUs) and micro-accelerators. Through operator fusion, on-chip memory scheduling, and specialized edge compilation toolchains, we eliminate hardware memory bottlenecks and achieve ultra-low-latency on-device inference for robotics, wearables, and edge vision.

We develop edge-centric AIoT systems that unify on-device vision with IoT sensor networks for low-latency inference and autonomous field control. Deployed across smart farms and event venues, our hierarchical Edge Hubs process environmental telemetry and video streams locally, ensuring uninterrupted operations while performing real-time analytics like crop morphometry, yield forecasting, and crowd hazard detection.
Building on this infrastructure, we design machine learning pipelines for closed-loop actuation and proactive decision-making. By correlating multimodal sensor data with visual analytics, the system triggers automated interventions—such as microclimate regulation in agricultural facilities or crowd routing and emergency alerts at venue sites—monitored seamlessly through a synchronized cloud-edge hybrid dashboard.

XR Rehabilitation developed by MR(Mixed Reality) system
It is a simulation that creates and attaches a virtual prosthetic leg on amputee’s leg in real-time
XR Rehabilitation performs EMG Signal Processing through Raspberry Pi to measure the patient’s EMG signal and is transmitted to Unity through TCP communication
XR rehabilitation can support non-face-to-face online services and can support face-to-face experience and immersion without physical constraint

478 3D facial landmarks (468 facial and 10 eye landmarks) are extracted from 2D RGB images, so a single camera yields three-dimensional facial geometry.
Landmarks captured at different times differ in scale, rotation, and position and cannot be compared directly. They are registered first by scale matching, then by global registration for a coarse alignment, and finally by point-to-plane ICP initialized with the transformation matrix from that alignment.
After registration, distance symmetry and angular symmetry give a numerical measure of left-right symmetry in a static expression.
For dynamic expressions, the amount of landmark movement compares the two sides of the face at the same instant.

Improved Feature-based Gaze Estimation Using Self-Attention Module and Synthetic Eye Images
Estimates gaze by detecting eye region landmarks through a single eye image
Rich information by extracting more landmarks and including iris and eye edges, 9 similar to the existing feature-based methods
HRNet backbone network to learn representations of images at various resolutions
Self-attention module CBAM to obtain a refined feature map with better spatial information