Multi-Feature Embedding and Deep Classification for Elderly Activity Recognition

Human Activity Recognition (HAR) has many potential applications in the real world, including biometric user identification, elderly health monitoring, and governmental surveillance, and as a result, it is currently the subject of intensive study. HAR has emerged as a major focus in the fields of mobile and omnipresent computing because to the increasing prevalence of wearable sensors and the Internet of Things. In recent years, deep learning has been the dominant method for making inferences and solving problems in the HAR system. However, significant difficulties remain in applying HAR to issues in biometric user identification, whereby a wide range of human behaviors may be considered biometric features and utilized to distinguish individuals. Using deep learning models to recognize human behavior as its foundation, this study presents an original framework for multi-class wearable identification of users. Sensory data from the wearable devices' tri-axial gyroscopic and tri-axial accelerometers are used to learn more about users while they engage in a variety of activities. Further, the usefulness of the suggested framework was shown by a series of tests. According to the results, the Convolutional Neural Network's model had the most overall accuracy (91.77
^%
) across all users, while the Long Short-Term Memory machine learning model achieved the highest overall accuracy (92.43%). Regarding biometric user identification, both of them are good enough.