NICEIR: NLOS identification of CSI system based on enriched features and improved random forest algorithm

Non-line-of-sight (NLOS) identification is a major challenge for reliable WiFi-based sensing. Existing NLOS identification methods commonly encounter limited statistical features, rely on pre-designed identification thresholds, and suffer from large computational complexity. To address these challenges, this paper presents a novel method called NLOS identification of channel state information system based on enriched features and improved random forest algorithm (NICEIR). A novel statistical feature is designed to enrich the feature set for recognition by incorporating amplitude variance and phase difference variance across all subcarriers. Furthermore, an improved random forest algorithm with C4.5 architecture is proposed to train the combined features and accurately discriminate NLOS conditions from incoming undetermined signals, eliminating the need for identification thresholds. Two pruning methods are developed for feature selection to reduce complexity and prevent overfitting. Experimental results across three different scenarios demonstrate the method's high recognition accuracy, low computational complexity, and broad applicability.