上肢建模運用於關節遮擋問題以及軌跡相似度比較

Upper Extremity Modeling For Joint Occlusion and Trajectory Similarity Comparison

王俊麒(Jyun-Ci Wang)

指導教授 ： 周承復

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103819

摘要

在中風復健系統中，準確性、可靠性和遮擋是三個重要的關注面向。由於現在多數方法著重在解決前兩者準確性和可靠性的問題，但是遮擋也會對系統判別造成影響，因此我們提出融合異質性的感測器，RGB-D相機以及穿戴式裝置，搭配機器學習的方式建構出上肢模型，可在畫面出現遮擋問題時由模型預測出被遮擋的關節位置，以利系統進行正確的判定，並滿足準確性、可靠性和遮擋的三個問題。

有了一系列復健動作的關節位置軌跡，為了矯正患者進行復健的動作，我們希望能將患者以及標準的復健動作進行比對，提供患者能更有效率進行復健。我們提出signature descriptor來代表整段軌跡，因此只要對signature進行比對就能得出兩段軌跡的相似程度，對患者這次的復健動作打出分數。

關鍵字

機器學習 ； 神經網路

並列摘要

In stroke rehabilitation system, there are three important focuses: accuracy, reliability, and occlusion. Most recent research focus on high accuracy and reliability, but occlusion problem also play a big role in system judgement. Therefore we propose to fuse heterogeneous data sensors: RGB-D camera and wearable devices, and construct an upper extremity model with machine learning-based method. When occlusion problem happens, the model can predict the occluded joint location, for the system to function correctly. At the same time satisfy the three focuses: accuracy, reliability, and occlusion.

With a sequence of the trajectory of the rehabilitation motion joint location. In order to correct the patient's rehabilitation motion, we hope to compare the patient's motion with the standard motion, and make the patient rehabilitate more efficiently. We propose a signature descriptor to represent the whole trajectory. Therefore we only need to compare signature and we can get the similarity of the two trajectories, and give a score of the patient's rehabilitation motion.

並列關鍵字

Machine Learning ； Neural Network

參考文獻

[1] A. Paszke, S. Gross, F. A. Lerer, J. Bradbury, G.Chanan, T. Killeen, Z. Lin, N.Gimelshein, and L.Antiga, et al. Pytorch: An imperative style, highperformance deep learning library. Advances in neural information processing systems, 2019.

[2] C. Chang, S. Wang, and C. Wang. Exploiting moving objects: Multirobot simultaneous localization and tracking. IEEE Transactions on Automation Science and Engineering, 2016.

[3] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Skeletonbased action recognition with convolutional neural networks. IEEE International Conference on Multimedia and Expo Workshops, 2019.

[4] J. Chen, C. Wang, E. Wu, and C. Chou. Simultaneous heterogeneous sensor localization, joint tracking, and upper extremity modeling for stroke rehabilitation. IEEE Systems Journal, page 1–12, 2020.

[5] D. Carroll. A quantitative test of upper extremity function. Journal of chronic diseases, 1965.