利用加速度計與全卷積類神經網路之步態參數萃取

Accelerometer Only Gait Parameter Extraction With All Convolutional Neural Network

陳昶宏(Chang-Hong Chen)

指導教授 ： 張添烜

國立交通大學/電機學院/電子工程系所/碩士(2018年)

摘要

近年以來，步態參數萃取相關應用逐漸興盛，不論是在醫療照護領域、老年人照護、運動訓練、日常行走監控，因此如果有適當的步態參數當作指標，在許多領域上都能提供許多實際的應用，並真正幫助人類。
早期有許多研究需要透過攝影機系統配合取得步態參數，卻受限於室內的環境設定，而也有許多傳統的方法採用6軸感測器，透過二次積分後並搭配許多繁雜的校正演算法，進而推得步態參數，然而除了過程中繁多的校正演算法，藉由位置進一步推得之參數也會有一定的誤差累積。
本論文提出全卷積類神經網路(all-convolutional-layer CNN)模型，僅用三軸加速規之加速度值便能訓練並建立步態參數之模型，取得每一步之步態資訊包含步長、步高，以及總移動距離。所提出的全卷積層的CNN模型針對此步態參數應用，提出增加準確度與縮減複雜度的訓練方法，減少了大量的訓練參數以及模型複雜度，除了計算量低之外，更少去傳統二次積分法所面臨的複雜的校正過程，此方法相較早期攝影機系統侷限於室內環境，於室內/室外環境皆適用。
在準確度提升上，由於資料量在模型訓練時扮演相當重要的角色，本論文提出兩種資料增強的方法，增加20倍的資料讓模型去學習，而兩種資料增強應用於步高、步長之模型測試時，準確度也有所提升，而在訓練過程採用最佳模型選擇，除了避開測試時較大的錯誤率，一方面也提升了準確度。
在複雜度縮減上，所提出的模型簡化也進一步大大減少了參數量與模型複雜度，所提出之模型除了在高採樣率500Hz下做到步態參數萃取，在低採樣率50Hz也能成功萃取步態參數，除了電腦上能做到步態參數萃取，本論文成功將所提出的模型實現於智慧型手機APP上，即時呈現結果，證明此方法於實際應用之可行性。
本論文採用交叉驗證來測試所提出的模型之效果與實用性。測試資料包含了2154步的走路資料。交叉驗證下，步長的萃取可以達到平均4.29%的百分比誤差，而平均誤差為3.56公分，在步高的部分，可以達到平均百分比誤差11.65%，而其平均誤差為0.63公分，而距離估算在1689公尺的總距離下可以達到0.1%的百分比誤差。

關鍵字

步態參數 ； 加速度計 ； 全卷積類神經網路

並列摘要

Application of gait parameter extraction blow up in the last few years. For example, medical care, elderly care, sports training, and monitor of daily walking. With the appropriate gait parameter as an index, gait parameter extraction can provide many practical implementations in our daily life and help human.
There were some work using the camera system for the research on the gait parameter extraction in the past. The method has the disadvantage of restricting to indoor environmental settings. There are also many double- integration methods using the data from 6-axis sensor. The gait parameter are computed through the double integration and complex algorithm of calibration. In addition to complex calibration procedures, there are also some error accumulations in the computation of gait parameter, which is inferring from the position after double integration.
This thesis proposed a CNN model with a single tri-axial accelerometer to do the gait parameter extraction, including stride length, stride height, and the total moving distance. Proposed model is the all-convolutional-layer CNN model for gait parameter extraction. This thesis proposed training method for improving accuracy and reducing complexity, which has greatly reduced the training parameter and the model complexity. The relatively small filter size results in a simpler model with low computational work. Besides, it can be used in both indoor and outdoor environment instead of the complex calibration in the tradition double-integral method.
Considering the improving of accuracy, data amount plays an important role in the training of the model. This thesis proposes two data augmentation methods, increasing twenty times data amount for model training. The testing results of proposed model on stride length and stride height with data augmentations both have better result.
Considering the reducing of model complexity, proposed method on model simplification has greatly reduced the training parameters and model complexity. Proposed model on gait parameter extraction is done at high sampling rate 500Hz and lower sampling rate 50Hz. Besides, this thesis has successfully done the implementation on android phone and the presenting of the result in real time, which implies the feasibility in practical application.
To verify the model on the result of gait parameter extraction, this thesis uses the cross validation. There are total 2154 steps walking data. The proposed work can reach the average percentage error 4.29% with the average error 3.56cm on stride length extraction. With respect to stride height extraction, proposed model can reach the average percentage error 11.65% with the average error 0.63cm. Moreover, the percentage error of total walking distance estimation is 0.1% with 1689 meters.

並列關鍵字

Gait parameter ； Accelerometer ； All Convolutional Neural Network

參考文獻

[1] N.-H. Ho, P. H. Truong, and G.-M. Jeong, “Step-detection and adaptive step-length estimation for pedestrian dead-reckoning at various walking speeds using a smartphone,” Sensors, vol. 16, no. 9, p. 1423, 2016.

[2] Harvey WeinBerg, “AN-602: Using the ADXL202 in Pedometer and Personal Navigation Applications,” Analog Devices, Tech. Rep., 2002.

[3] J. W. Kim and C. Park, “A Step , Stride and Heading Determination for the Pedestrian Navigation System,” Technology, vol. 3, no. 1, pp. 273–279, 2005.

[4] Q. Tian, Z. Salcic, K. I.-K. Wang, and Y. Pan, “A multi-mode dead reckoning system for pedestrian tracking using smartphones,” IEEE Sensors J., vol. 16, no. 7, pp. 2079–2093, Apr. 2016.

[5] W.-Y. Shih, L.-Y. Chen, and K.-C. Lan, “Estimating walking distance with a smart phone,” in Parallel Architectures, Algorithms and Programming (PAAP), 2012 Fifth International Symposium on. IEEE, 2012, pp.166–171.