Personalized Activity Recognition with Deep Triplet Embeddings.

Area of Science:

• Computer Science
• Machine Learning
• Biomedical Engineering

Background:

• Supervised learning for human activity recognition faces challenges due to user-specific data variations.
• This heterogeneity leads to suboptimal performance for certain individuals in activity recognition systems.

Purpose of the Study:

• To develop a personalized human activity recognition approach using deep feature representation.
• To investigate the effectiveness of novel loss functions, including a subject triplet loss, for improving model robustness.

Main Methods:

• A convolutional neural network (CNN) was employed for deep feature extraction.
• Categorical cross-entropy and triplet loss functions were used for training, with a novel subject triplet loss proposed.
• Performance was evaluated on MHEALTH, WISDM, and SPAR datasets, assessing classification accuracy and out-of-distribution detection.

Main Results:

• The proposed triplet loss algorithm achieved an average classification accuracy of 96.7%, significantly outperforming the baseline CNN's 87.5%.
• Personalized algorithms demonstrated superior robustness against inter-subject variability.
• Enhanced performance was observed in both classification and out-of-distribution activity detection tasks.

Conclusions:

• Personalized deep learning, particularly with novel triplet loss functions, effectively mitigates inter-subject variability in human activity recognition.
• The proposed approach offers a more robust and accurate solution for inertial sensor-based activity recognition systems.
• This method shows promise for real-world applications requiring reliable human activity monitoring.