近年來，透過 Wi-Fi 訊號進行人類活動識別（Human Activity Recognition, HAR）的研究逐漸受到關注。相較於傳統的影像處理方法或穿戴式感測技術，Wi-Fi 擁有非侵入性、隱私保護性高以及無需穿戴設備等特點，使其特別適用於智慧家庭、醫療照護、長者生活監控與安防系統等應用場域，展現出高度的實用潛力與研究價值。在技術發展與應用需求不斷交織推動下，Wi-Fi-based HAR 系統正逐漸由實驗室走向商品化，例如 Origin Wireless 已與 ASUS 合作推出基於 Wi-Fi 的居家動作偵測產品，顯示此技術具備相當的產業發展潛力。 然而，Wi-Fi 訊號本身極易受到各種環境因素的干擾，例如牆面反射、傢俱遮蔽、人員移動與天候變化等，這些條件的改變會顯著影響無線訊號在空間中的傳遞特性，進而導致不同場域下所獲得的通道特性（channel characteristics）呈現出高度變異性。這類空間與時間上的不穩定性，使得從原始 Wi-Fi 資料中萃取出能夠泛化至不同環境的行為特徵變得極具挑戰性。因此，發展具備環境適應性與結構穩定性的訊號處理與特徵建模策略，已成為 Wi-Fi 感知研究中一項關鍵課題。 為了應對上述挑戰，本研究提出一種以統計方法為基礎的特徵濾析流程，藉由建模 CSI（Channel State Information）訊號的時頻變異結構，篩選出與人體活動具有高度關聯性的訊號分量，藉此達到強化行為相關性與抑制背景干擾的目的。相較於傳統以全域特徵為主的識別方式，本方法更強調動態變化中的局部特徵與其跨時序的一致性，以提升模型在面對環境轉換時的辨識準確性與泛化能力。此外，隨著 Wi-Fi 6（802.11ax）等新世代無線通訊協議的普及，本研究亦採用支援 Wi-Fi 6 的 Intel AX210 無線網卡進行資料擷取，該裝置可提供更高的 CSI 時頻解析度與更穩定的訊號品質，為 Wi-Fi 感知技術未來的進一步應用與擴展提供了更堅實的基礎。 ;Human activity recognition (HAR) using Wi-Fi signals has gained increasing attention due to its non-intrusiveness, high privacy, and lack of reliance on wearable devices. These advantages make Wi-Fi sensing promising for applications in smart homes, healthcare, and elderly monitoring. Recent collaborations, such as that between Origin Wireless and ASUS, have demonstrated the potential of commercializing Wi-Fi-based HAR systems. Despite these advancements, Wi-Fi signals are highly sensitive to environmental changes, including wall reflections, occlusions, and spatial variations. These disturbances lead to unstable channel state information (CSI), making it difficult to extract robust and generalizable activity features across different environments. Addressing this challenge requires signal processing techniques that can suppress irrelevant variations and highlight motion-relevant patterns. In this work, we propose a statistical feature refinement framework that models the time-frequency structure of CSI signals to isolate components closely related to human motion. Our approach focuses on dynamic local features and their temporal consistency, enabling improved cross-environment recognition performance. Furthermore, we leverage Wi-Fi 6 (802.11ax) technology and the Intel AX210 wireless card to collect high-resolution and stable CSI data, laying a foundation for scalable and reliable Wi-Fi-based sensing.
