The NeuroFeel Project
A final year software engineering project exploring cross-dataset emotion recognition using physiological signals
BEng (Hons) Software Engineering
This project was developed as part of the Final Year Project (FYP) for the BEng (Hons) Software Engineering degree program at IIT, affiliated with University of Westminster.
Project Details
NeuroFeel investigates the challenges of cross-dataset emotion recognition using physiological signals, presenting two innovative frameworks: one for personalized emotion recognition using minimal calibration data, and another employing advanced domain adaptation techniques to enhance model generalization from controlled laboratory environments to real-world scenarios.
NeuroFeel is a research-oriented project dedicated to enhancing emotion recognition using physiological signals. It focuses specifically on developing frameworks that handle both personalized emotion detection and cross-domain generalization, thereby ensuring consistent performance across different individuals and settings.
The project comprises two main frameworks: a personalized emotion recognition framework built on the WESAD dataset, which utilizes transfer learning and adaptive ensemble modeling for effective subject-specific emotion classification; and a cross-dataset emotion recognition framework designed to bridge the gap between controlled laboratory (WESAD) and naturalistic real-world (K-EmoCon) environments through advanced domain adaptation methods such as CORAL, Subspace Alignment, and ensemble adaptation techniques.
Key Project Components
WESAD Personalization
Adaptive model selection for enhanced personalized emotion recognition
Cross-Dataset Framework
Domain adaptation techniques for real-world generalizability
Interactive Demo
Visual comparison of both frameworks with real-time results
Sensor Discrepancies
The laboratory WESAD dataset used high-quality sensors at 700Hz, while K-EmoCon used consumer-grade wearables at 4Hz. I developed specialized normalization techniques to address this fundamental quality gap.
Emotion Taxonomy Mapping
Bridging discrete emotions (WESAD) with dimensional valence-arousal ratings (K-EmoCon) required creating a data-driven mapping between these fundamentally different emotion representation approaches.
Performance Evaluation
Developing fair comparison metrics across different datasets and emotion models presented unique challenges. I implemented balanced evaluation methods to ensure accurate assessment of cross-dataset performance.
WESAD Dataset
Used for academic research only
Schmidt et al. (2018), ICMI
K-EmoCon Dataset
Requires access approval
Park et al. (2020), Scientific Data
Experience NeuroFeel in Action
Try the interactive demo to see how the frameworks perform across different datasets and scenarios.