Analysis of Effective Subcarrier Bandwidths for Different Object Detections

This paper presents a CSI-based human detection system with commodity sensors: M5Stack. Moreover, we analyze the proposed system with SHAP and clarify important features for the final prediction because the system usually predict whether human exists or not with nonlinear machine learning and we cannot interpret relations between input features and the prediction. CSI represents WiFi propagation conditions and we can detect environmental changes. To measure CSI, we usually need specific devices to extract CSI information from WiFi packets. However, we construct a CSI measurement system at a low cost using commodity sensors: M5Stack. We measure CSI with our proposed system and detect human presence with Gradient Boosting Decision Tree, which is a nonlinear ensemble machine learning, according to observed CSI amplitudes. The proposed system is superior to a decision tree, which is a simpler machine learning algorithm, from the viewpoint of prediction accuracy. Speaking concretely, the proposed system achieves approximately 71.3% and is superior to decision tree because the proposed system can capture more complex characteristics for prediction. From the analysis results, we find contributions to the prediction are different for each feature in CSI.