Researcher
Incheon
Seoul, South Korea
From Signal Fidelity to Robotic Intent: Low-Channel EEG Restoration for Motor-Imagery BCI
1
Comparative EEG Channel Restoration for Low-Channel BCI
This project investigates whether a reduced motor-area EEG montage can be transformed into a common full-channel representation while preserving information relevant to motor-imagery decoding. Native low-channel input, interpolation, deterministic autoencoders, and conditional generative models were evaluated under the same leakage-controlled inter-session protocol with matched TCFormer classifiers. The results showed that restoration can recover part of the information lost through electrode reduction, while the spectral-spatial autoencoder provided the strongest overall restored-input decoding performance.
2
Intent Decoding and ROS2-Based End-to-End Command Interface
Restoration quality was evaluated across waveform, spectral, covariance, CSP-feature, and downstream decoding metrics rather than reconstruction error alone. The rankings differed across these criteria: methods that reproduced certain signal statistics more accurately did not necessarily provide the most useful representations for motor-imagery classification. This revealed a clear distinction between signal fidelity and task-relevant information, while computational benchmarking further showed that deterministic restoration offered a more practical accuracy–efficiency balance than iterative diffusion-based inference.
3
ROS2 Integration and LiDAR-Safe Simulated Actuation
The common EEG representation was integrated with a ROS2 pipeline linking motor-intent decoding to simulated assistive robot control. TCFormer predictions were converted into planar velocity commands, while a separate LiDAR-based safety filter modified or stopped motion according to the surrounding environment before commands reached the simulated wheelchair. This architecture separates learned intent estimation from deterministic safety control and demonstrates software-level continuity from prepared EEG trials through inference, safety filtering, and simulated actuation.
All photos are taken and edited by me



