Projects

Vision-language models for bite measurement and eating-rate feedback

NIH-funded DIBS project, URI. 2026–present.

Eating-rate feedback needs bite measurements that arrive during the meal, and rules that still behave sensibly when those measurements are wrong. I adapted Qwen2.5-VL with LoRA and stacked temporal attention to detect bites, chewing, or neither from a single profile-view camera, using only past frames, and then studied how its errors propagate into feedback.

• Evaluated on 29 laboratory participants, each held out in turn, with model selection restricted to inner-validation participants. Mean bite-event F1 was 0.887, onset error 382 ms, and 60-second bite-rate error 0.43 bites/min, versus 1.63 for a no-video baseline.
• Removing temporal attention lowered event F1 by 0.034. I also compared CLIP, VideoMAE, and DINOv2 encoders and a linear head against the language decoder.
• Detection accuracy and feedback agreement are separate questions: injected errors changed the meal-long rule’s fast/not-fast label in about 10% of evaluated seconds but changed reminders far more, with disagreement above 25%.

Results are from recorded meals and simulation; prospective intervention evaluation is future work. Submitted to IEEE JBHI.

Tripolar concentric ring electrode EEG

URI. 2026–present.

A methods and pilot comparison of saline-soaked felt, conductive gel, and paste electrode constructions on resting-state and visual-evoked-potential recordings. I led the Python analysis pipeline: harmonized preprocessing, scale-invariant cross-setup metrics (alpha SNR, eyes-open/eyes-closed reactivity, spectrogram similarity), preprocessing sensitivity analyses, and per-subject VEP diagnostics. First-author manuscript in preparation. Code

EEG and motion capture for movement decoding

GNN_MOCAP_TNLAB project, URI. 2025–present.

A synchronized acquisition platform combining 32-channel g.tec EEG, OptiTrack Motive motion capture, and StretchSense glove kinematics with hardware TTL triggers, plus experimenter and participant interfaces, counterbalanced finger and grasp tasks, impedance monitoring, and trial-level timing and signal-quality checks. Pilot analyses use low-frequency EEG, lagged regression, held-out trials, and artifact and null controls to examine finger kinematics.

Computer vision for neuromuscular assessment

URI. 2025–present.

Markerless hand-pose analysis and vision-based assessment methods for socially assistive robotics and neuromuscular rehabilitation research.

Earlier work: wearable sensor validation

URI. 2023.

Evaluated IMU-based knee-angle measurements against optical motion capture and analyzed calibration and sensor-placement effects, which led to two papers (see Publications).

• YCB object segmentation: fine-tuned YOLOv26n-seg for object segmentation in images and video, to support robotic perception.
• EEG motor-intention classification: filter bank common spatial patterns, mutual-information feature selection, and linear discriminant analysis for grasp-type classification.
• 3D hand pose estimation from RGB: a custom CNN compared with a pretrained ResNet50 for regression of 21 hand joints on FreiHAND, with cross-validation.