Abstract
Wi-Fi CSI signal shows the state of a communication link and obstacle in the communication path changes its observable criteria. This physical phenomenon can be leveraged to gather human activity through applying a model and learning change patterns. The paper presents an attempt to improve a model-based human sensing method by signal preprocessing and upgrading deep learning model performance. Then we apply the softmax regression algorithm for activity classification. An experiment shows that our proposed approach makes it possible to achieve better accuracy (15% higher) and test time (0.79 seconds faster) for all activities when compared with pattern-based approach.