臺灣博碩士論文加值系統

本研究利用攝影機在跑步機上取得的平面影像，將其轉換為骨架節點資訊後，僅使用骨架節點資訊，來達到身份辨識的目的。
研究的出發點是希望能捨棄掉完整的影像資料，以便在有隱私性較強的場所中（如居家環境、寢室等，有可能錄製到人物衣衫不整等情況）也可以僅透過讀取骨架節點資訊來做到人物辨識的功能。
研究總共分成三個階段，以及一個原地踏步影像的實驗。第一個階段我們在跑步機上搜集四個人、四種不同角度的跑步資料，找出最適合分辨人物的特徵關節點並建立模型，使其可以分別在知道角度的前提下辨識出人物身份。第二個階段我們嘗試透過訓練一個轉換角度的模型，讓不同角度的影像，也可以轉換成單一角度，以此，可以僅使用一種角度的身份辨識模型來辨識身份。第三個階段，我們透過機器學習的方法，讓系統可以辨識目前的影像屬於何種角度。
透過以上三個階段的研究，我們就可以轉換不同角度的跑步資料，並透過一個身份辨識模型，辨識出人物身份。
而在原地踏步實驗方面，我們使用不同的骨架節點預測框架，測試是否在不同的運動情形、不同的框架下，仍能辨識出該影像是由何種角度錄製的。

This article is based on analyzing the data of skeleton key points, that are transformed by 2D videos of people running on treadmills. The goal is to use these data from skeleton key points to identify a person and a position they are in.
The reason why we discard the video and use the skeleton data is because there are some privacy issues. If we put a camera at home, it might record people naked walking in the room, but they don’t want this to be saved. To avoid this, my research is going to use only skeleton data to identify a person.
This research is separated in three parts. In the first part, I recorded videos of people running on treadmills using four different camera positions. I tried to find the key points which can tell the differences between people and used these key points to build models that can identify a person. In the second part, I tried to transform one position’s data to another, so that I can identify a person in different camera position by using only one model. In the last part, I tried to train a model to show which position is recorded.
After these researches, I have a method to show different person running on the treadmill in different positions.

摘要 i
Abstract ii
誌謝 iii
目錄 iv
圖目錄 vi
表目錄 vii
第一章 緒論 1
1.1 研究背景 1
1.2 研究動機 1
1.3 研究目標 1
1.4 研究流程 2
第二章 實驗設計與設備 4
2.1 研究資料組 4
2.1.1 實驗一：以四種角度在跑步機上跑步 4
2.1.2 實驗二：以五種角度原地踏步 5
2.2 研究設備 5
第三章 研究方法 6
3.1 跑步機上跑步 6
3.1.1 影像轉換至骨架節點 6
3.1.2 資料前處理 7
3.1.3 切割出每步起始時間 8
3.1.4 觀察步間資料 9
3.1.5 統一資料長度 10
3.1.6 建立模型及特徵選取 11
3.1.7 轉換角度 14
3.1.8 辨識角度 14
3.1.9 KNN (K Nearest Neighbor) 15
3.2 原地踏步 16
3.2.1 影像轉換至骨架節點 16
3.2.2 分群 16
第四章 實驗結果 17
4.1 跑步機上跑步 17
4.1.1 特徵選取結果 17
4.1.2 各角度模型預測結果 19
4.1.3 轉換角度結果 22
4.1.4 辨識角度結果 25
4.2 原地踏步 26
4.2.1 辨識角度結果 26
第五章 結論與建議 27
5.1 結論 27
5.2 未來使用情境 27
5.3 未來研究方向 27
參考文獻 28

[1] Y. Lu, A. Fleury, J. Boonaert, S. Lecoeuche, and S. Ambellouis. Online person identification and new person discovery using appearance features. In 2015 IEEE International Conference on Evolving and Adaptive Intelligent Systems (EAIS), pages 1–8, Dec 2015.
[2] N. Radha, A. Shahina, and A. Nayeemulla Khan. A person identification system combining recognition of face and lip-read passwords. In 2015 International Conference on Computing and Network Communications (CoCoNet), pages 882–885, Dec 2015.
[3] Vasileios Belagiannis, Andrew Zisserman. Recurrent Human Pose Estimation. In 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017), pages 468 – 475, May 2017.
[4] Zhe Cao and Gines Hidalgo, Tomas Simon, Shih-En Wei, Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. arXiv preprint arXiv:1812.08008. Dec 2018.
[5] Zhe Cao, Tomas Simon, Shih-En Wei, Yaser Sheikh. Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. arXiv:1611.08050. Apr 2017.
[6] Tomas Simon, Hanbyul Joo, Iain Matthews, Yaser Sheikh. Hand Keypoint Detection in Single Images using Multiview Bootstrapping. arXiv:1704.07809. Apr 2017.
[7] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, Yaser Sheikh. Convolutional pose machines. arXiv:1602.00134v4. Apr 2016.
[8] Yong Du, Wei Wang, Liang Wang. Hierarchical Recurrent Neural Network for Skeleton Based Action Recognition. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1110-1118, 2015