Event-based Human Action Recognition: Rehabilitation Monitoring Dataset and Models

Human Action Recognition (HAR) enables interactive rehabilitation monitoring for patients and has wide range of use cases in healthcare. Event-based vision sensors have become of interest for such deployment scenarios as they offer high temporal resolution and operate under varying illumination, while inherently preserving patient privacy by capturing only motion-triggered changes in the scene. Although there is growing interest in event-based HAR, current datasets dominantly cover general activities and lack focus on movements specific to the rehabilitation. In this work, we introduce the first open-source event-driven dataset tailored for rehabilitation exercises, study three models with distinct features covering ResNeXt-50, a CNN-LSTM based and a Fully CNN based model, and investigate their ability to learn from this dataset. Our results provide key insights into the applicability of event-based sensing for privacy-preserving and real-time rehabilitation monitoring.