本論文以FMCW毫米波雷達相位訊號進行非接觸式身分辨 識研究。相較於傳統影像式或接觸式身分辨識方法，毫米波雷 達不需取得人體影像，亦不需使用者配戴任何裝置，因此具有隱 私保護、非接觸式感測與不受光照條件影響等優點。本研究使用 IWR6843ISK-ODS 毫米波雷達搭配DCA1000EVM資料擷取板取得 人體目標之原始ADC訊號，並透過資料重組、方向估測、波束成 型、Range FFT、目標距離單元選取、相位擷取與相位解纏等處理 流程，取得可反映人體胸腔微動之雷達相位訊號。 在特徵建構與模型設計方面，本研究分別採用人工設計相位 特徵與深度時頻特徵進行身分辨識。首先，由雷達相位時間序列 中萃取20維候選人工設計特徵，經PCA特徵負荷量分析後選取 其中14維，並使用ECOC-SVM進行分類。其次，將相位訊號轉 換為時頻圖，作為ResNet18深度學習模型之輸入，以學習雷達相 位訊號於時間與頻率分布中的深度特徵。為提升CNN對關鍵時頻 區域之特徵表示能力，本研究進一步比較SE、Non-local與CBAM 三種注意力機制。此外，本文亦針對完整頻段、呼吸頻段與心跳 頻段建立ResNet18-CBAM模型，以分析不同頻段時頻特徵對身分 辨識之影響。其中，完整頻段設定為0.02–5.00Hz，呼吸頻段設定 為0.10–0.50 Hz，心跳頻段設定為0.80–2.00Hz。 為整合不同特徵來源之互補資訊，本文將ECOC-SVM類別 分數，以及完整頻段、呼吸頻段與心跳頻段之深度時頻特徵，分 別建立為四個異質專家分支，並透過混合專家模型（Mixture-of Experts， MoE）進行融合。模型利用Top-2稀疏門控機制，根據輸 入樣本特性自適應選取門控權重最高之兩個專家分支，並融合其 特徵資訊，產生最終身分辨識結果。 實驗結果顯示，ECOC-SVM使用14維人工設計相位特徵，可 達到78.90% 之整體準確率。在深度學習模型中，PureResNet18 Fullband 之整體準確率為64.40%，加入SE與Non-local 後皆提升 至77.60%，而加入CBAM後則提升至78.90%。此結果顯示，注 意力機制能改善CNN對雷達時頻圖之特徵表示能力，其中同時結 合通道注意力與空間注意力之CBAM，於完整頻段ResNet18模型 中具有最佳整體辨識表現。 進一步透過MoE融合四個專家分支後，整體準確率提升至 82.40%，相較於最佳單一模型提升3.50個百分點。 此外，本研究亦建立即時身分辨識系統，整合初始資料緩衝、 週期性辨識、30秒多數決與人體晃動偵測機制，以驗證所提出方 法於即時應用情境下之可行性。綜合實驗結果可知，FMCW毫米 波雷達相位訊號具備應用於非接觸式身分辨識之可行性；結合人 工設計相位特徵、深度時頻特徵與多專家融合架構，則可有效提 升雷達身分辨識系統之整體準確率，並改善部分低辨識率類別之 辨識表現。;This thesis investigates non-contact human identity recognition using phase signals acquired by an FMCW millimeter-wave radar. Com pared with conventional image-based or contact-based identity recognition methods， millimeter-wave radar does not require human images or wearable devices. It therefore offers advantages in privacy preservation， non-contact sensing， and robustness to lighting conditions. In this study， an IWR6843ISK-ODS millimeter-wave radar and a DCA1000EVM data capture board are used to acquire raw ADC signals reflected from human targets. Through data reshaping， direction estimation， beamforming， Range FFT， target range-bin selection， phase extraction， and phase unwrapping， radar phase signals that reflect chest micro-motion are obtained. For feature construction and model design， this study employs both handcrafted phase features and deep time-frequency features for identity recognition. First， 20 candidate handcrafted phase features are extracted from the radar phase time series. Based on an analysis of PCA loadings， 14 features are selected and used as inputs to an ECOC-SVM classifier. Second， the phase signals are converted into spectrograms and used as inputs to ResNet18-based deep learning models to learn temporal and spectral representations from the radar phase signals. To enhance the CNN feature representation in informative time-frequency regions， three attention mechanisms， namely SE， Non-local， and CBAM， are further compared. In addition， ResNet18-CBAM models are constructed for the fullband， respiration-band， and heartbeat-band spectrograms to analyze the effects of different frequency-band features on identity recognition. The fullband， respiration band， and heartbeat band are set to 0.02–5.00 Hz， 0.10–0.50 Hz， and 0.80–2.00 Hz， respectively. To integrate complementary information from different feature sources， the ECOC-SVM class scores and the deep time-frequency fea- tures from the fullband， respiration-band， and heartbeat-band models are organized into four heterogeneous expert branches. These expert branches are fused using a Mixture-of-Experts (MoE) model with a Top 2 sparse gating mechanism. For each input sample， the model adaptively selects the two expert branches with the highest gating weights and combines their feature representations to produce the final identity prediction. The experimental results show that the ECOC-SVM using 14- dimensional handcrafted phase features achieves an overall accuracy of 78.90%. Among the deep learning models， Pure ResNet18-Fullband achieves an overall accuracy of 64.40%， whereas ResNet18-SE-Fullband and ResNet18-Nonlocal-Fullband both achieve 77.60%. ResNet18 CBAM-Fullband achieves the highest accuracy among the fullband CNN models at 78.90%. These results indicate that attention mechanisms im- prove the feature representation capability of CNNs for radar spectrograms. Among the fullband ResNet18 models， CBAM achieves the best overall recognition performance by combining channel attention and spatial attention. After the four expert branches are fused through the MoE model， the overall accuracy increases to 82.40%， which is 3.50 percentage points higher than that of the best standalone model. This study also implements a real-time identity recognition system that integrates initial data buffering， periodic recognition， 30-second majority voting， and human-motion detection. When human motion is detected， the current recognition cycle is skipped， and the system displays a recognition-unavailable status. The implemented system is used to verify the feasibility of the proposed method in a real-time application scenario. Overall， the experimental results demonstrate the feasibility of using FMCW millimeter-wave radar phase signals for non-contact human identity recognition. The integration of handcrafted phase features， deep time frequency features， and a multi-expert fusion architecture improves the overall recognition accuracy and enhances the recognition performance of several classes with low recall.
