Location-Aware Ambient Sound Model Adaptation for On-Device Human Activity Recognition in Living Spaces

초록

Human activity recognition (HAR) using sound-based approaches provides a non-intrusive and practical solution for detecting and classifying human activities in real-world environments. However, deploying HAR systems across diverse environments presents challenges due to domain shift and variations in environmental acoustics. Traditional HAR models trained on pre-collected datasets often fail to generalize to new settings, leading to performance degradation and misclassification errors. This paper proposes an adaptive ambient sound model optimization system to enhance the robustness and adaptability of HAR in edge computing environments. The system can dynamically customize classification labels based on the installation environment, ensuring flexibility across different locations. To mitigate domain shift, edge-based transfer learning fine-tunes a pre-trained model using locally collected data, improving classification accuracy across varying acoustic conditions. Additionally, a pseudo-labeling mechanism continuously optimizes the model by iteratively refining predictions from high-confidence unlabeled data, enabling long term adaptability without extensive manual annotations. To validate the proposed approach, we implemented the system and conducted experiments using real-world sound data from multiple residential environments. Experimental results demonstrate that the system effectively adapts to different locations while maintaining high classification accuracy, real-time inference, and a short training time, making it suitable for practical edge deployment.

키워드