High-Level Features for Human Activity Recognition and Modeling

High-Level Features (HLF) are a novel way of describing and processing human activities. Each feature captures an interpretable aspect of activities, and a unique combination of HLFs defines an activity. In this article, we propose and evaluate a concise set of six HLFs on and across the CSL-SHARE and UniMiB SHAR datasets, showing that HLFs can be successfully extracted with machine learning methods and that in this HLF-space activities can be classified across datasets as well as in imbalanced and few-shot learning settings. Furthermore, we illustrate how classification errors can be attributed to specific HLF extractors. In person-independent 5-fold cross-validations, the proposed HLFs are extracted from 68% up to 99% balanced accuracy, and activity classification achieves 89.7% (CSL-SHARE) and 67.3% (UniMiB SHAR) accuracy. Imbalanced and few-shot learning results are promising, with the latter converging quickly. In a person-dependent evaluation across both datasets, 78% accuracy is achieved. These results demonstrate the possibilities and advantages of the proposed high-level, extensible, and interpretable feature space.