Human activity recognition (HAR) based on inertial measurement units (IMUs) embedded in wearable devices has gained increasing relevance in healthcare, wellness, and fitness monitoring. However, most existing classification methods assume a closed-set setting, where all activity classes need to be defined during training, which limits their applicability in real-world environments where unseen or unexpected activities are present. To overcome this limitation, we adopt an open-set recognition (OSR) framework that requires minimal changes to the HAR classifiers traditionally employed for this purpose. We also provide an extensive empirical evaluation based on a leave-one-activity-out validation protocol applied to two datasets with IMU signals acquired from smart eyewear: a proprietary dataset and the publicly available UCA-EHAR dataset. A lightweight one-dimensional convolutional neural network was trained to classify six-axis IMU data across common activities. We assess open-set HAR performance using several methods requiring limited computational overhead and operating in the logit space, including maximum logit, Gaussian Mixture Models, Kernel Density Estimation, OpenMax, and Nearest Neighbor Distance Ratio. Robust identification of unknown activities was achieved, with area under the ROC curve > 0.8. These findings highlight the potential of low-complexity open-set approaches for real-time HAR on resource-constrained wearable platforms, supporting the development of adaptive and reliable sensor-based recognition systems for real-world use.
Cortese et al. (Fri,) studied this question.