Activity Recognition from Channel State Information for Few-Sampled Scenarios

Human Activity Recognition (HAR) using channel state information (CSI) from wireless systems has been a field in progress for the last few years. In particular, amplitudes of WiFi-based CSI have been widely used for HAR in indoor scenarios. In contrast, CSI phases are less exploited due to higher sensitivity to timing errors in channel estimation. This paper further explores a recent phase processing method to obtain environmental information from CSI and improves classification using a Convolutional Neural Network in few-sampled scenarios. For this purpose, a new dataset has been generated with measurements of three receivers on different days and with different people, for activity recognition and people counting. In addition, we incorporate a transfer learning strategy using fine-tuning along with a data fusion model from different receivers to improve accuracy metrics by training the model and validating it on data taken on different days. Promising results with accuracies higher than 80 % have been obtained with transfer learning between days using 5 % of data for fine-tuning.