Feature Extraction with Scalogram-Based CNN Activations and Activity Recognition in Elderly Individuals with Transfer Learning

The growing elderly population requires healthcare professionals to utilize the Human Activity Recognition (HAR) system accurately and objectively. Wearable sensor technologies offer objective measurement-based HAR methods as a solution. In this study, a signal processing-based deep learning approach was used to assess the daily activities of elderly individuals. The scalogram images, which are the time-frequency representation of Raw signals obtained from three axial accelerometers in the HAR70+ dataset, were used as input to the common Convolutional Neural Networks (CNN) with a transfer learning approach. Then, the activations of the global average pooling layer of the CNN networks were used as features of the machine learning algorithms. ResNet18 was the most successful CNN network with 87.37% F1-Score value, while the combination of Xception and LDA was the most successful one with 88.37% F1-Score value during testing.