Open-Set Recognition of Human Activities from Head-Mounted Inertial Sensor.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Human Activity Recognition (HAR) using Inertial Measurement Units (IMUs) is crucial for healthcare and fitness.
• Existing HAR methods struggle with real-world scenarios due to the presence of unseen activities (closed-set limitation).

Purpose of the Study:

• To adapt HAR classifiers for real-world applicability by implementing an open-set recognition (OSR) framework.
• To evaluate the effectiveness of OSR in identifying unknown activities using smart eyewear IMU data.

Main Methods:

• A lightweight 1D convolutional neural network was trained on six-axis IMU data.
• An extensive evaluation used a leave-one-activity-out protocol on two datasets (proprietary and UCA-EHAR).
• Open-set HAR performance was assessed using logit-space methods like GMM, KDE, OpenMax, and NNdr.

Main Results:

• Robust identification of unknown activities was achieved with an Area Under the ROC Curve (AUC) greater than 0.8.
• The OSR framework demonstrated effectiveness with minimal computational overhead.
• The study validated findings on both a proprietary and a public dataset (UCA-EHAR).

Conclusions:

• Low-complexity OSR approaches hold significant potential for real-time HAR on resource-constrained wearable devices.
• This research supports the development of adaptive and reliable sensor-based recognition systems for practical applications.
• The findings pave the way for more robust HAR systems capable of handling novel activities.