利用相機擷取影像的非接觸式心跳訊號提取技術

Abstract

近年來，由於非接觸和低成本的優點，使用相機進行光體積量測成為一門熱門的研究領域。此類技術能夠在不接觸皮膚的情況下量測出心臟的生理資訊，如：心跳頻率、心率變異。然而，過去發表的研究成果對於非靜止的受測者，量測心跳頻率的準確度相當有限，特別是當劇烈動作所產生的干擾其頻率落在跟心跳頻率相同的頻帶內。另外，先前的研究對於用相機量測心率變異的表現也是相當有限，因為心率變異是由心跳間隔換算出來的各種統計數值，而心跳間隔本身對於雜訊和干擾特別敏感，即使量測目標靜止不動，也很難取得準確的心跳間隔進而換算成準確的心率變異。本論文針對上述兩種心臟生理資訊，分別提出對應的非接觸式方法解決在特定應用上遭遇的問題。第一個是針對運動中的心跳頻率量測提出的運動穩健演算法，第二個是針對日常生活使用手機或平板的情境下準確量測使用者心率變異的方法。我們提出的心跳頻率量測方法能夠移除運動造成的干擾並準確量測到在運動器材(如：跑步機)上劇烈運動的受試者心跳，實驗證實我們的演算法能夠把平均準確率從71% 改善到94%。另一方面，我們提出的心率變異量測方法能夠在不同測試條件下，得到相當接近於接觸式裝置所得到的心跳間隔所換算出來的心率變異數值。根據實驗結果，對於所有測試影片，我們的演算法所得到的心率變異指標的平均絕對值誤差僅有4.33 (毫秒)。

Recently, the camera-based photoplethysmography (PPG) became a popular research field due to the non-contact and low-cost properties. These techniques are able to measure the cardiac physiology, e.g., heart rate, heart rate variability (HRV) without contacting the skin of subject. However, the previous works showed limited success to measure the heart rate when subjects are not stationary, especially when the interference rendered by both vigorous motions and heart beat are in the same frequency band. Besides, the previous works showed limited success for estimating the HRV since the HRV is computed by the beat-to-beat intervals (also known as R-R intervals, RRI) which are sensitive to noise and artifacts. Even when the subject is static, it is still difficult to obtain reliable RRI for computing accurate HRV. This dissertation proposed two contactless algorithms corresponding to two different applications: the first one is motion-robust heart rate estimation for exercising subject; the second one is accurate HRV measurement for monitoring the HRV of the smart phone or tablet users in daily-life scenario. The proposed heart rate estimation can remove the motion artifacts even when subject is exercising on some fitness machines (e.g. treadmill). Experiments show that our algorithm significantly improved the averaged estimation accuracy from 71% to 94% correct. The proposed HRV estimation method is able to obtain reliable HRV metrics which are close to the ones measured by contact device under different conditions. As shown in experiments, the averaged error of HRV metrics obtained by our method is only 4.33 (in the unit of millisecond) for all the video clips.