Event-based Vision for Privacy-Preserving Patient Monitoring

Abstract

This study proposes and is developing a novel approach for de-identifying and classifying the actions of bedridden patients using an RGB-DVS hybrid sensor system.

We are first leveraging an RGB camera to identify the patient’s initial state. A DVS generation model, which has been pre-trained on DVS data, is then being used to create de-identified data.

Subsequently, we are developing a model to estimate the patient’s current posture from this de-identified data, which will result in an algorithm that can accurately predict the patient’s real-time state.

Key Features

• Privacy-First Design: De-identification through DVS data
• Hybrid Sensor System: RGB-DVS integration
• Real-time Monitoring: Continuous patient state prediction
• Healthcare Innovation: Addressing privacy concerns in medical settings

Research Pipeline

1. Initial State Detection: RGB camera for patient identification
2. De-identification: Pre-trained DVS generation model
3. Posture Estimation: Deep learning from de-identified data
4. Real-time Prediction: Continuous patient state monitoring

Applications

• Hospital patient monitoring
• Elderly care facilities
• Home healthcare systems
• Privacy-compliant medical AI