步態辨識與行走運動強度估測系統之研究

Abstract

心血管疾病患者常常需要經由醫師來確定患者的安全可運動強度與設計運動處方；肥胖者需要透過自主管理來控制體重。對於具有以上需求的使用者，發展一套可長時間監控的個人化運動強度紀錄系統，有其存在之必要性。 本文提出一個可行的個人化步態辨識與行走運動強度估測系統。在步態辨識部分，本文使用αβ濾波器得到精確的運動姿態，並將該姿態經過經驗模態分解(Empirical Mode Decomposition，EMD)濾除雜訊後提取傅利葉能量譜，進行線性鑑別式分析(Linear Discriminant Analysis，LDA)訓練與辨識。當辨識為走路狀態時，即進行行走運動強度估測。本文同時探討走路時物理慣性做功與運動強度的相關性。使用二次逼近與經驗模態分解(Empirical Mode Decomposition，EMD)之殘餘函數濾除加速規本身的基線飄移對於積分的嚴重影響，得到準確的瞬時速度，進而提升運動強度計算的準確性。 本文針對5男5女共10位受試者，測試四種步態(上樓、下樓、走路、跑步)之個人步態辨識在30%訓練樣本時，剩餘70%作驗證之平均準確率可達90%以上。每位受試者並進行10次不同行走速度的測試，在步數計算之平均準確率可達95%，並改善Y. Kurihara計算運動強度的方式，其迴歸方程式之相關性可由原本的0.55提升至0.81，證明本研究所採用之方法確實可提高運動強度估測之準確性。最後實作於智慧型手機，並於PC端開發使用者介面，協助了解運動中的姿態變化與演算法驗證，日後可以進一步達到遠端監控、雲端儲存、自主管理等目的。

Cardiovascular disease patients need suggestions from doctors for an acceptable level of exercise intensity and exercise prescription. And over-weight patients control their weight by self-management. Developing a monitoring system which can record personalized exercise intensity is necessary for above patients. This thesis proposes a feasible system for personalized gait recognition and walking exercise intensity estimation. On analyzing gait recognition, this thesis uses α-β filters to obtain better athletic attitudes, and further uses Empirical Mode Decomposition (EMD) to filter noise of athletic attitude to acquire Fourier Transform energy spectrum. Thus, the Linear Discriminant Analysis (LDA) can apply to this energy spectrum for training and recognition. When the motion is recognized as walking, the walking exercise intensity is estimated. This thesis also discusses the correlation between inertia work and exercise intensity by using residual function of Empirical Mode Decomposition (EMD) and quadratic approximation to filter the baseline shift of acceleration sensor to reduce the serious integral effect. And further we are can derive better exercise intensity and instantaneous speed. This thesis uses measured 10 subjects including 5 males and 5 females to recognize four types of gait from upstairs, downstairs, walking, and running. For gait recognition, 30% of collected raw data is used for training samples, and recognition rate of verifying the 70% data can reach more than 90%. After applying our method to ten different walking speeds from each subject, we found that step calculation shows 95% accuracy, and Y.Kurihara’s exercise intensity method can be enhanced, and the regression equation correlation can be increased from 0.55 to 0.81. These results prove our method can improve exercise intensity estimation. The proposed method has been implemented on smart phones and graphic user interface on personal computers to help understanding the change of athletic stance and to verify the proposed algorithm for further application on remote monitoring, cloud computing, and self-management.