A Framework for Human Activity Recognition Application for Therapeutic Purposes

The study of Human Activity Recognition (HAR) involves identifying and categorizing human activities through movements detected by wearable devices, sensors, or videos. HAR has been adopted in various domains, including physical therapy, where patients’ biomechanical abilities are monitored. This study proposes a framework for a human activity recognition-based solution to automate parts of the biomechanical assessment and monitor patients’ activities to assist physiotherapists and provides insight into metrics describing the progression of treatment. The proposed framework uses a visual programming interface to enable physical therapists to describe their specific data set and intended outputs using an intuitive graphical interface. An experiment was conducted to evaluate the performance of common recurrent deep learning architectures used in HAR, in addition to the investigation into the impact of a convolutional layer on their performance. Results indicate that CNN-LSTM is capable of 99.2% accuracy on the MHealth data set, indicating that it is an effective architecture for accurately classifying human activities.