Interpretable High-Level Features for Human Activity Recognition
by , , ,
Abstract:
This paper introduces and evaluates a novel way of processing human activities based on unique combinations of interpretable categorical high-level features with applications to classification, few-shot learning, as well as cross-dataset and cross-sensor comparison, combination, and analysis. Feature extraction is considered as a classification problem and solved with Hidden Markov Models making the feature space easily extensible. The feature extraction is person-independently evaluated on the CSL-SHARE and UniMiB SHAR datasets and achieves balanced accuracies up from 96.1% on CSL-SHARE and up to 91.1% on UniMiB SHAR. Furthermore, classification experiments on the separate and combined datasets achieve 85% (CSL-SHARE), 65% (UniMiB SHAR), and 74% (combined) accuracy. The few-shot learning experiments show potential with low errors in feature extraction but require further work for good activity classification. Remarkable is the possibility to attribute errors and indicate optimization areas easily. These experiments demonstrate the potential and possibilities of the proposed method and the high-level, extensible, and interpretable feature space.
Reference:
Interpretable High-Level Features for Human Activity Recognition (Yale Hartmann, Hui Liu, Steffen Lahrberg, Tanja Schultz), In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies - Volume 4: BIOSIGNALS, 2022.
Bibtex Entry:
@inproceedings{hartmann2022highlevelfeature,
  title = {Interpretable High-Level Features for Human Activity Recognition},
  author = {Hartmann, Yale and Liu, Hui and Lahrberg, Steffen and Schultz, Tanja},
  booktitle = {Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies - Volume 4: BIOSIGNALS},
  pages = {40-49},
  year = {2022},
  isbn = {978-989-758-552-4},
  issn = {2184-4305},
  doi = {10.5220/0010840500003123},
  abstract = {This paper introduces and evaluates a novel way of processing human activities based on unique combinations of interpretable categorical high-level features with applications to classification, few-shot learning, as well as cross-dataset and cross-sensor comparison, combination, and analysis. Feature extraction is considered as a classification problem and solved with Hidden Markov Models making the feature space easily extensible. The feature extraction is person-independently evaluated on the CSL-SHARE and UniMiB SHAR datasets and achieves balanced accuracies up from 96.1% on CSL-SHARE and up to 91.1% on UniMiB SHAR. Furthermore, classification experiments on the separate and combined datasets achieve 85% (CSL-SHARE), 65% (UniMiB SHAR), and 74% (combined) accuracy. The few-shot learning experiments show potential with low errors in feature extraction but require further work for good activity classification. Remarkable is the possibility to attribute errors and indicate optimization areas easily. These experiments demonstrate the potential and possibilities of the proposed method and the high-level, extensible, and interpretable feature space.},
  url = {https://www.csl.uni-bremen.de/cms/images/documents/publications/HartmannLiuLahrbergSchultz_Biosignals2022.pdf}
}