Human activity recognition (HAR) systems often involve vision-based sensors or wearable devices, which can be intrusive and raise privacy concerns. They also require dedicated hardware. This paper proposes a method for HAR based on Wi-Fi Channel State Information (CSI) data that moves toward explainable model decisions. This method enables non-invasive, device-free monitoring without the need for additional infrastructure. Using the publicly available NTU-Fi-HAR dataset, CSI signals corresponding to six activities, walking, running, boxing, falling, arm rotation, and floor cleaning were transformed into images and processed using Machine Learning techniques. Several feature extraction filters, classification models and hyperparameter configurations were evaluated. The best baseline performance was obtained using the ColorLayout filter and Random Forest model, achieving a test per packet accuracy of 0.73. Additionally, an occlusion-based explainability analysis was developed to highlight the image regions most relevant to the model’s decision. This analysis showed that the JpegCoefficient filter offers a better trade-off between performance and interpretability. Finally, a hierarchical majority-rule post-processing algorithm has been introduced to refine the output probabilities of the selected Machine Learning model. Two optimisation strategies were evaluated: treating the signals from the three receiver antennas independently versus aggregating their multi-antenna information. The overall aggregated accuracy was enhanced to 0.87 and 0.90, respectively.
Toward an Explainable Machine Learning Method for Wi-Fi CSI-Based Human Activity Recognition
Poce, Sara
;Varriano, Giulia;Carissimo, Chiara;Cerro, Gianni;Mercaldo, Francesco;Santone, Antonella
2026-01-01
Abstract
Human activity recognition (HAR) systems often involve vision-based sensors or wearable devices, which can be intrusive and raise privacy concerns. They also require dedicated hardware. This paper proposes a method for HAR based on Wi-Fi Channel State Information (CSI) data that moves toward explainable model decisions. This method enables non-invasive, device-free monitoring without the need for additional infrastructure. Using the publicly available NTU-Fi-HAR dataset, CSI signals corresponding to six activities, walking, running, boxing, falling, arm rotation, and floor cleaning were transformed into images and processed using Machine Learning techniques. Several feature extraction filters, classification models and hyperparameter configurations were evaluated. The best baseline performance was obtained using the ColorLayout filter and Random Forest model, achieving a test per packet accuracy of 0.73. Additionally, an occlusion-based explainability analysis was developed to highlight the image regions most relevant to the model’s decision. This analysis showed that the JpegCoefficient filter offers a better trade-off between performance and interpretability. Finally, a hierarchical majority-rule post-processing algorithm has been introduced to refine the output probabilities of the selected Machine Learning model. Two optimisation strategies were evaluated: treating the signals from the three receiver antennas independently versus aggregating their multi-antenna information. The overall aggregated accuracy was enhanced to 0.87 and 0.90, respectively.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


