SeWi: A Framework Enhancing CSI-Based Human Activity Recognition

With the widespread application of WiFi devices, utilizing Channel State Information (CSI) for human activity recognition has garnered significant attention. However, the reliability of WiFi sensing signals is often compromised due to the operation of most WiFi devices in complex electromagnetic environments. Consequently, the accuracy of current recognition systems often falls short of expectations, impeding the practical application of WiFi sensing. We discover that segmenting and reassembling data can yield features from multiple perspectives, thereby enhancing recognition accuracy. In this paper, we propose a versatile segmentation component capable of employing various segmentation methods. Based on this, we develop the SeWi learning framework to improve the accuracy of the original basic model. In our experiments, we analyze the sliding window segmentation method and utilize four different models as the basic models for SeWi. We also analyze the effective range of hyperparameters for this segmentation method. The results indicate that SeWi exhibits varying degrees of improvement for different models. Of particular note, using ResNet18 as the basic model for SeWi, the accuracy achieved on the Widar and UT-HAR public datasets is 74.8% and 98.7%, respectively. To the best of our knowledge, these results represent an increase of 3.1% and 0.6%, respectively, compared to the best results reported in recent studies.