Hand Washing Gesture Recognition Using Synthetic Dataset

Area of Science:

• Computer Vision
• Machine Learning
• Public Health

Background:

• Effective hand hygiene is crucial in healthcare and food industries.
• Autonomous systems are needed to evaluate hand washing compliance.
• Limited availability of comprehensive hand washing datasets hinders development.

Purpose of the Study:

• To address the data scarcity challenge for hand hygiene recognition.
• To present an open synthetic dataset for training machine learning models.
• To evaluate the performance of different neural network models using this synthetic data.

Main Methods:

• Created a synthetic dataset with 96,000 frames of 3D computer-generated hand washing gestures.
• Included RGB, depth, and hand mask images across eight gestures, four characters, and four environments.
• Trained and evaluated four models (Inception-V3, Yolo-8n, Yolo-8n segmentation, PointNet) on the dataset.

Main Results:

• Models trained on synthetic data achieved notable classification accuracies on real-world data.
• Yolo-8n segmentation reached 79.3% accuracy, Yolo-8n achieved 76.3%, and Inception-V3 reached 56.9%.
• Demonstrated the utility of synthetic data for improving hand washing gesture recognition.

Conclusions:

• Synthetic data is effective for training machine learning models for hand washing gesture recognition.
• The developed dataset can facilitate the creation of autonomous hand hygiene monitoring systems.
• This approach offers a viable solution to the data limitations in the field.