[Paper Review] Privacy-Preserving Action Recognition for Smart Hospitals using Low-Resolution Depth Images

Edward Chou, Matthew Tan+5 more|arXiv (Cornell University)|Nov 25, 2018

Medical Imaging and Analysis19 citations

TL;DR

This paper proposes a privacy-preserving action recognition framework for smart hospitals using low-resolution depth images to minimize patient surveillance concerns. By downsampling depth frames to 14×14 or 56×56 and enhancing them via a privately-trained DCSCN super-resolution model, the method maintains high accuracy in hand-hygiene monitoring (98.24%) and ICU activity logging (72.4% accuracy), demonstrating that privacy-preserving vision systems can retain sufficient utility for critical healthcare tasks.

ABSTRACT

Computer-vision hospital systems can greatly assist healthcare workers and improve medical facility treatment, but often face patient resistance due to the perceived intrusiveness and violation of privacy associated with visual surveillance. We downsample video frames to extremely low resolutions to degrade private information from surveillance videos. We measure the amount of activity-recognition information retained in low resolution depth images, and also apply a privately-trained DCSCN super-resolution model to enhance the utility of our images. We implement our techniques with two actual healthcare-surveillance scenarios, hand-hygiene compliance and ICU activity-logging, and show that our privacy-preserving techniques preserve enough information for realistic healthcare tasks.

Motivation & Objective

• Address patient privacy concerns in hospital surveillance by minimizing visual information in depth images.
• Maintain sufficient action recognition utility for critical healthcare tasks despite aggressive downsampling.
• Develop a framework that combines low-resolution input with private super-resolution enhancement to balance privacy and performance.
• Demonstrate the feasibility of deploying non-intrusive vision systems in real-world healthcare settings like ICUs and hygiene compliance monitoring.
• Enable future integration of RGB-based systems with privacy guarantees using similar low-resolution and enhancement techniques.

Proposed method

• Downsample original 224×224 depth images to 14×14 or 56×56 using bicubic interpolation to reduce privacy-sensitive information.
• Train a privately-trained DCSCN super-resolution model on a separate, privacy-safe dataset to enhance low-resolution depth images.
• Use a ResNet-50 model for hand-hygiene recognition and a ResNet-18 model for ICU activity logging, both fine-tuned on downsampled and enhanced frames.
• Apply data augmentation with random transformations to balance class imbalances in the hand-hygiene dataset.
• Employ a two-stage pipeline: low-resolution input → super-resolution enhancement → action recognition inference.
• Ensure privacy by training the super-resolution model on non-patient data, avoiding exposure to sensitive medical or identity information.

Experimental results

Research questions

• RQ1 To what extent does downsampling depth images to 14×14 or 56×56 preserve action recognition utility for healthcare tasks?
• RQ2 Can a privately-trained super-resolution model enhance low-resolution depth images without reintroducing privacy risks?
• RQ3 Does the combination of downsampling and super-resolution improve or degrade performance compared to full-resolution inputs?
• RQ4 How do privacy-preserving downsampling and enhancement compare to existing methods in terms of accuracy and AUC on real healthcare action recognition tasks?
• RQ5 Can low-resolution depth-based systems achieve performance comparable to or better than full-resolution systems in specific healthcare applications?

Key findings

• The 56×56 resolution with DCSCN enhancement achieved 98.24% test accuracy in hand-hygiene recognition, surpassing the original 224×224 baseline of 94.5%.
• Even the 14×14 resolution with DCSCN enhancement achieved 95.87% accuracy, outperforming the original resolution in the hand-hygiene task.
• For ICU activity logging, the 56×56 resolution with DCSCN achieved 72.4% test accuracy, showing minimal degradation compared to the 68.8% baseline at full resolution.
• The AUC for ICU class 4 ('get out of chair') improved from 0.724 (14×14, no enhancement) to 0.725 (with DCSCN), indicating enhanced discrimination despite low resolution.
• Visual inspection confirmed that 14×14 depth frames are visually unrecognizable for human identity, facial features, or sensitive activities, supporting strong privacy guarantees.
• Surprisingly, downsampling alone improved performance in some cases, likely due to regularization effects reducing overfitting on the hand-hygiene dataset.