FusedNet: A Fusion of Time Series and Imaging Based Human Activity Recognition Using ResNet

Human Activity Recognition (HAR) is one of the active research in ubiquitous computing that covers various fields such as robotics, surveillance. Sensor-based HAR with advanced deep learning models suffers from the overfitting and vanishing gradient problem because of one-dimensional time series data. Therefore, the proposed FusedNet framework addresses the aforementioned issue using a fusion of 1D and a two-dimensional image obtained by Gramian Angular Field (GAF). Initially, 1D and GAF-based images are processed using convolutional neural network (CNN) and residual CNN network, respectively. Next, the fusion of 1D and 2D information is performed for HAR prediction. As a result, the issues are resolved with the aid of residual connection and fusion of 1D information. Additionally, in the experimental analysis employing the two public datasets UCI-HAR and WISDM, the proposed model achieves accuracy of 94.33% and 98.53%, respectively.