LSTM is the widely used DL model for HAR—including in IoT-based HAR—because its memory capacity can deal better with time series data (such as HAR data) than other models, including CNN. The LSTM implementation of HAR can support transfer learning and is suitable for resource-constrained IoT devices. Generally, FER relies on image processing, and the CNN is the best model for image processing. Therefore, we implement use case two (FER) using a CNN model. In Chapter 3, Image Recognition in IoT, we presented an overview of two popular implementations of the CNN (such as incentive V3 and Mobilenets) and their corresponding transfer learning. In the following paragraphs, we briefly present an overview of the baseline LSTM.
LSTM is an extension of RNNs. Many variants of LSTM are proposed, and they follow...
