Hands-On Deep Learning for IoT

Deep learning models for HAR and emotion detection

Generally, human activity recognition systems use accelerometer and gyroscope signals, which are time series data. Sometimes, the recognition process uses a combination of time series and spatial data. In this context, Recurrent Neural Network (RNN) and LSTM are potential candidates for the former type of signals because of their capability to incorporate temporal features of input during evolution. On the other hand, CNNs are good for spatial aspects of accelerometer and gyroscope signals. Hence, a combination or hybrid of CNNs and LSTMs/RNNs is ideal for the former type of signals. We will use an LSTM model for the HAR use case as it can address the temporal aspects of human activities. ...