Generalizing location-centric variations to enhance contactless human activity recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Contactless Human Activity Recognition (HAR) is vital for smart healthcare and elderly care.
• Location-specific Channel State Information (CSI) variations hinder HAR model generalization across different environments.
• Existing HAR models struggle to perform effectively in unseen physical locations due to data heterogeneity and non-IID distributions.

Purpose of the Study:

• To develop a novel federated learning (FL) algorithm for robust contactless Human Activity Recognition (HAR).
• To address and mitigate location-induced data disparities and non-Independent and Identically Distributed (non-IID) data distributions in HAR.
• To improve the generalization capability of HAR models across diverse and unseen physical locations.

Main Methods:

• Proposed a novel federated learning (FL) algorithm named Federated Weighted Averaging for HAR (Fed-WAHAR).
• Implemented a dynamic weighting approach based on local models' accuracy to enhance global model performance.
• Evaluated Fed-WAHAR using metrics such as accuracy, precision, recall, F1 score, and convergence analysis.

Main Results:

• Fed-WAHAR achieved 85% accuracy in recognizing human activities across different locations.
• The algorithm effectively mitigated location-induced disparities and non-IID data distributions.
• Demonstrated reduced convergence time and improved global model classification accuracy.

Conclusions:

• Fed-WAHAR significantly enhances the robustness and generalizability of contactless HAR models.
• The proposed FL approach enables effective cross-domain HAR inference in unseen environments.
• This method holds promise for real-time monitoring in smart healthcare and elderly care applications.