本研究開發一套結合FMCW毫米波雷達與跨注意力特徵融合 之即時非接觸式心率估測系統，旨在提升非完全靜止情境下心率 估測之準確度與穩健性。系統透過DCA1000EVM擷取雷達原始 IQ 資料，經資料重組、距離向快速傅立葉轉換（RangeFFT）、目 標鎖定、空間濾波與相位擷取後，從連續雷達資料中萃取與胸腔 微小位移相關之生理特徵，以進行即時心率估測。 為提升不同訊號品質下的估測穩健性，本研究利用目標距離 單元（rangebin）及其鄰近距離單元建立多維度生理相位特徵，包 含原始相位、相位差分與濾波相位，作為深度學習模型之輸入。 除傳統頻譜峰值心率估測方法外，本文建立心率直接迴歸模型作 為深度學習比較基準，並利用Gaussian心跳事件估測模型產生時 間域心跳事件序列，作為融合模型之輔助資訊。在此基礎上，本 研究進一步提出Cross-Attention 雙路特徵融合模型，整合雷達生理 相位特徵與Gaussian 心跳事件序列，使模型學習兩種特徵來源之 時間關聯，以強化心跳相關特徵之辨識能力。 實測結果顯示，在靜止量測場景下，各深度學習模型之均方 根誤差（rootmeansquare error， RMSE）皆低於1BPM。在各方法最 終設定下，坐姿自然動作場景中，傳統FFT方法受到動作所造成 之相位干擾影響，其RMSE上升至14.98BPM；相較之下，所提出 之5RBCross-Attention 融合模型可將RMSE降至3.43BPM，平均絕 對誤差（meanabsolute error， MAE）降至 2.18 BPM，且 Pass±3 BPM 達76.8%，顯示其可有效提升非完全靜止情境下之心率估測穩定 性。此外，將最終模型直接應用於四位未參與模型訓練、驗證及 參數選擇之新增受測者時，在不進行重新訓練或個人化調整的情 況下，靜止量測與坐姿自然動作場景之平均RMSE分別為1.35BPM 與4.34 BPM，呈現初步的跨受測者適用性。本研究亦將雷達訊號 處理與深度學習模型整合至即時系統中，驗證其於連續雷達資 料流下進行即時心率推論之可行性。整體而言，FMCW雷達結合 Cross-Attention 特徵融合具有應用於非完全靜止情境下即時非接觸 式心率估測之潛力。;This study develops a real-time， non-contact heart rate estimation system that combines frequency-modulated continuous-wave (FMCW) millimeter-wave radar with Cross-Attention-based feature fusion， aim ing to improve the accuracy and robustness of heart rate estimation under non-static conditions. RawradarIQdataareacquiredbyaDCA1000EVM data acquisition module and then processed through data reconstruction， range fast Fourier transform (Range FFT)， target localization， spatial fil tering， and phase extraction. Physiological features associated with sub tle chest displacements are then extracted from continuous radar data streams for real-time heart rate estimation. To improve estimation robustness under varying signal-quality con ditions， multidimensional physiological phase features are constructed using the target range bin and its neighboring range bins. These features include the original phase， phase difference， and filtered phase， and are used as inputs to the deep learning models. In addition to the conven tional spectral peak-based heart rate estimation method， a direct heart rate regression model is established as a deep learning baseline. A Gaussian heartbeat event estimation model is also employed to generate a temporal heartbeat event sequence as auxiliary information for the fusion model. Based on this framework， a dual-branch Cross-Attention feature fusion model is further proposed to integrate the radar physiological phase fea tures with the Gaussian heartbeat event sequence， enabling the model to learn temporal relationships between the two feature sources and enhance the identification of heartbeat-related features. Experimental results show that， under static measurement condi tions， all deep learning models achieve root mean square error (RMSE) values below 1 BPM. Under the final configuration of each method， the conventional FFT-based method is susceptible to motion-induced phase interference in the seated natural-motion scenario， resulting in an RMSE of 14.98 BPM. Incontrast， theproposed5RBCross-Attentionfusionmodel reduces the RMSE to 3.43 BPM and the mean absolute error (MAE) to 2.18 BPM， while achieving a Pass±3 BPM rate of 76.8%， demonstrating im proved robustness of heart rate estimation under non-static conditions. In addition， when the final model is directly applied to four additional sub jects who were not involved in model training， validation， or parameter selection， without retraining or subject-specific adaptation， the average RMSE values are 1.35 BPM under static conditions and 4.34 BPM under seated natural-motionconditions， indicatingpreliminarycross-subjectap plicability. The radar signal-processing pipeline and deep learning mod els are also integrated into a real-time system， verifying the feasibility of real-time heart rate inference from continuous radar data streams. Over all， combining FMCW radar with Cross-Attention-based feature fusion shows promising potential for real-time， non-contact heart rate estima tion under non-static conditions.
