近年來，隨著遠距醫療（Telemedicine）與非接觸式生理技術的蓬勃發展，透過常規光學視訊攝影機連續量測人體心血管動力學已成為重要研究趨勢。其中，心衝擊圖（Ballistocardiogram， BCG）源於心室收縮射血對主動脈壁產生之微觀機械的反作用力，其物理本質為體表機械加速度（a(t)=\frac{d^2y(t)}{dt^2}），具備評估心臟收縮動能與動態心律的巨大潛力。 然而，傳統基於空間強度梯度之特徵追蹤技術（如 Kanade–Lucas–Tomasi， KLT 光流法），在提取微弱體表 BCG 時面臨三大嚴苛物理瓶頸：第一、人體平坦皮膚缺乏高對比紋理，容易引發結構張量矩陣奇異（\mathbit{Z}^T\mathbit{Z}\rightarrow0），導致相機熱雜訊被放大為隨機漫步漂移；第二、相機 8-bit 模數轉換器（ADC）之空間灰階量化截斷與像素鎖定效應（Pixel-Locking Effect），導致連續亞像素位移退化為鋸齒階梯狀波形；第三、將幾何位移轉換為體表加速度時，時間微分算子導致高頻雜訊功率譜密度隨頻率呈現幾何級數發散，產生劇烈數值脈衝並徹底吞沒生理 J 峰波形。 為了徹底克服上述物理限制，本研究提出一套基於「空間注意力引導」與「空間域複數局部相位」之抗體動非接觸式 BCG 加速度重構演算法。系統於受測者額頭黏貼高對比圓形黑色標記貼片並裁剪固定 128\times128 像素之感興趣區域（ROI）。演算法首先利用二維複數可操縱金字塔（Complex Steerable Pyramid， CSP）對影像進行 4 尺度與 4 方向之多解析度解構，將微動映射為空間解析信號之連續相角旋轉；接著，基於最佳線性不偏估計（BLUE）框架推導出振幅平方空間注意力加權機制（W(x，y)\propto A^2(x，y)），自動抑制低對比皮膚背景並聚焦於高訊噪比之貼片邊緣，藉由大數法則空間統計池化在微分前有效壓制微動底噪。在運動降維與特徵選擇方面，本研究將各定向子帶微動速度先經一階時間微分轉換為體表加速度，隨後於各尺度內部透過主成分分析（PCA）完成多方向運動解耦以提取代表性 BCG 加速度波形，結合動態窄帶諧噪比（Dynamic HNR）與四階統計峰度（Kurtosis）構建綜合訊號品質指標（SQI），自適應選出最佳生理尺度通道。最後，演算法利用三次樣條升頻插值（1000\mathrm{\ Hz}）與生理不反應期雙重 約束，精準定位 BCG 加速度 J 峰並提取拍對拍心搏間距（Inter-Beat Interval， IBI）。 本研究透過頻率域相位平移合成平台與各向同性 2D 高斯斑紋消融實驗，驗證了演算法在極端低照度與感測器熱雜訊下突破 8-bit 量化極限的數值穩定性。在 10位健康受試者之人體臨床驗證中，以同步接觸式第一導程心電圖（Lead-I ECG）之 RR 間期作為黃金標準：在靜止情境下，本方法提取之 IBI 展現極高的線性相關性（r=0.991）與決定係數（R^2=0.982），Bland-Altman 95% 一致性界限（LoA）顯著收窄至 [-18.9，19.2] ms，大幅優於傳統 KLT；在受試者進行微幅點頭與轉頭之動態體動干擾情境下，相較於 KLT 相關性之徹底崩潰（r=0.191），本方法仍展現強韌之動態追蹤效能（r=0.936，R^2=0.876）。本研究從數理上解決了KLT光流之微分發散與紋理依賴難題，為高精度非接觸式心血管功能監控與穿戴式替代方案提供了具備嚴謹理論基礎之工程解決架構。;Non-contact physiological monitoring using conventional optical cameras has emerged as a transformative paradigm in telemedicine and pervasive healthcare. Ballistocardiogram (BCG) signals， which originate from the mechanical recoil forces exerted on the vascular walls during systolic cardiac blood ejection， physically represent micro-scale bodily accelerations (a(t)=\frac{d^2y(t)}{dt^2}) that offer profound insights into myocardial contractility and dynamic cardiac intervals. However， conventional video-based motion tracking methods based on spatial intensity gradients (e.g.， the Kanade–Lucas–Tomasi， KLT optical flow tracker) face three severe physical and numerical constraints when extracting micro-scale bodily movements: first， the lack of prominent geometric textures on smooth human facial skin causes ill-conditioned structure tensors (\mathbit{Z}^T\mathbit{Z}\rightarrow0)， heavily amplifying camera thermal noise into random-walk drifts; second， the discrete spatial quantization of 8-bit analog-to-digital converters (ADCs) and pixel-locking effects degrade continuous sub-pixel displacements into staircase artifacts; third， when converting geometric displacement to acceleration， time differentiation acts as a high-pass operator， causing high-frequency noise power spectral density to explode and generate impulse spikes that completely obscure the underlying physiological J-peaks. To overcome these fundamental limitations， this study proposes a motion-robust non-contact BCG acceleration reconstruction framework guided by spatial attention and complex local phase optical flow. The proposed system employs a high-contrast circular black marker patch placed on the subject’s forehead to define a consistent 128\times128 pixel region of interest (ROI). Using a 2D Complex Steerable Pyramid (CSP) across 4 scales and 4 orientations， bodily micro-motions are translated into continuous sub-pixel phase rotations via the spatial analytic signal representation. An amplitude-squared spatial attention weighting mechanism (W(x，y)\propto A^2(x，y)) is theoretically derived under the Best Linear Unbiased Estimator (BLUE) framework to suppress textureless skin noise and focus computation on high-SNR marker boundaries， drastically lowering the front-end noise floor prior to differentiation via the Law of Large Numbers. Furthermore， directional micro-velocities are converted to bodily surface accelerations via first-order time differentiation. Intra-scale 4-orientation Principal Component Analysis (PCA) is then performed to decouple multi-axis recoil trajectories and project dominant acceleration signals. Subsequently， an adaptive Signal Quality Index (SQI)—integrating dynamic narrow-band Harmonic-to-Noise Ratio (HNR) with statistical Kurtosis—is formulated to dynamically select the optimal spatial scale channel. Finally， sub-frame cubic spline upsampling (1000\mathrm{\ Hz}) coupled with physiological refractory period constraints is utilized for accurate J-peak localization and beat-to-beat Inter-Beat Interval (IBI) extraction. The numerical stability and resolution beyond 8-bit quantization limits were validated through synthetic Fourier phase-shift simulations and isotropic 2D Gaussian speckle ablation studies under severe noise and low-illumination conditions. In clinical human trials involving 10 subjects synchronized with Lead-I Electrocardiogram (ECG) ground truth， the proposed method achieved superior beat-to-beat IBI tracking fidelity at rest (r=0.991， R^2=0.982)， with the Bland-Altman 95% Limits of Agreement (LoA) significantly narrowed to [-18.9，19.2] ms. Under non-voluntary head motion disturbances， while KLT tracking collapsed entirely (r=0.191)， the proposed framework preserved high tracking robustness (r=0.936， R^2=0.876). This research resolves the fundamental derivative noise divergence and texture dependence of intensity-based optical flow， presenting a mathematically grounded and motion-resilient engineering solution for non-contact cardiovascular monitoring.
