一種人體姿態命令辨識及其身份識別的強化式方法之研究

An Enhanced Method on Human Gesture-Based Command and Command Maker Recognition

張育瑞(Yu-Jui Chang)

指導教授 ： 丁英智

國立虎尾科技大學/電機資訊學院/電機工程系碩士班/碩士(2016年)

摘要

本論文提出一套基於姿態命令辨識的使用者身份識別系統，此系統建構於KINECT的基礎上，並利用3D人體骨架座標發展具有強化姿態辨識的改良式特徵值與強化身份識別的融合特徵值，以及用於使用者合法性確認的模型學習偏移量。本論文使用三種方法進行辨識研究，分別是主成份分析(Principal components analysis, PCA)、隱藏式馬可夫模型(hidden Markov model, HMM)與支撐向量機(Support vector machine, SVM)。
在強化式人體姿態辨識的研究上採用PCA演算法，主要使用自行開發之活動姿態偵測(gesture activity detection, GAD)來與PCA結合強化，此法有助於提升資料的完整性與有效性，爾後更會融入PCA的重心距離決策流程。實驗結果證實結合GAD的方法皆優於傳統PCA的辨識率。
在強化式身份辨識的研究上採用HMM演算法，而使用的融合特徵值是由肢體角度、相鄰關節長度與肢體面積所組成，因其獨有的時序特性能讓使用者在操作動作的過程依照個別習慣而被加以區分。而學習的方式會依照學習資料的多寡各別發展，在學習資料量足夠的情況下採用模型內差機器學習，每次的學習過程都會得到模型的偏移量，在資料量缺乏時便使用融合HMM姿態辨識之極少量資料學習法，而經過篩選的資料還會經過改良式的內差學習法加以學習，以及經過Type-2 Fuzzy的調整致使學習效率達到最佳化。經實驗結果顯示有經過篩選的少資料學習方法性能優於傳統的模型學習，不僅使用的資料量大幅減少，辨識準確率亦獲得大幅提升。
於使用者合法性確認使用SVM演算法，特徵值則使用HMM的模型內差機器學習所得到的學習偏移量，利用每位使用者的動作習慣不一的特性，即使操作相同的動作也會得到不一樣的偏移結果，作者依照此特性將學習偏移量用於合性與非法的分類，每次的學習將會使得模型更加接近使用者的動作習慣。依照實驗的結果顯示，HMM模型學習偏移量確實有效於合法性的研究。
最後本論文的實用層面建構於物聯網的命令控制應用，結合作者提出的方法模擬居家的控制應用與安全篩檢，另外會加入例外資料庫於姿態辨識的方法，防止無意義之動作也會觸發控制。

關鍵字

KINECT ； 物聯網 ； 模型學習偏移量 ； PCA ； HMM ； SVM ； GAD

並列摘要

This thesis uses Microsoft’s KINECT device to develop a user recognition system that is on the basis of gesture commands. This paper uses three different types of features for recognition, modified features for enhanced gesture recognition, converged features for enhanced user identification, and model learning offset features for checking the legitimacy of users. All these three designed features in this thesis are extended from 3D human skeleton coordinates. For recognition processing, the author uses three different kinds of methods, principal components analysis (PCA) for enhanced user identification, hidden Markov model (HMM) for enhanced user identification, and support vector machine (SVM) for user legitimacy verification.
In PCA enhanced gesture recognition, a new algorithm called gesture activity detection (GAD), is proposed to improve the process of gesture data collection by elevating the effectiveness of the obtained data and then extracting the meaningful data segment. Furthermore, GAD is integrated into the decision-making process in PCA gesture recognition to improve the recognition rate. Experimental results show the superiority of the GAD method.
In the development of enhanced user identification using HMM, the converged feature consists of limb angles, lengths of adjacent joints and limb areas. The unique time-varying characteristic of HMM can be used for classifying each users by his own habits of motions. The learning method design for enhanced user identification by HMM considers the quantity of data. When the learning data is sufficient, we can get model offsets after every learning process, and such rich gesture data are finely referenced for model interpolations. In the case of lack of learning data, user recognition by gesture commands uses the strategy of learning data filtering. Experimental result shows that enhanced user identification using HMM with recognition system learning proposed in this thesis is better than traditional user identification by HMM without any learning tasks.
For user recognition to check the legitimacy of the user by using SVM, the offset of HMM model learning is treated as the feature. According to the different motion habits of all users, the offset feature will still have a fine classification effect even if they operate the same gesture. The offset characteristic will get closer to the user after every learning offset re-estimate. Experimental results show user recognition using SVM with the feature of the offset of HMM model learning is very effective.
Finally, presented methods in this thesis can be considered in a practical application internet of things (IOT). To simulate the device control using gesture commands and security screening of the command maker in a home environment, a system framework to implement the application system using proposed methods in this thesis has been planned in this thesis.

並列關鍵字

KINECT ； IOT ； model learning offset ； PCA ； HMM ； SVM ； GAD

參考文獻

1.R. Banerjee, A. Sinha and K. Chakravarty, “Gait Based People Identification System Using Multiple Switching Kinects,” International Conference on Intelligent Systems Design and Applications (ISDA), Bangi, pp. 182-187, 8-10 Dec, 2013.

5.A. Corradini and H. Gross, “Camera-based Gesture Recognition for Robot Control,” Proc. IEEE-INNS International Joint Conference, vol. 4, pp.133-138, 2000.

6.E. Rakun, M. Andriani, I. W. Wiprayoga, K. Danniswara and A. Tjandra, “Combining Depth Image and Skeleton Data from Kinect for Recognition Words in The Sign System for Indonesian Language,” Proc. International Conference on Advanced Computer Science and Information Systems, pp. 387-392, 2013.

7.M. A. Livingston, J. Sebastian, A. Zhuming and J. W. Decker, “Performance Measurements for The Microsoft Kinect Skeleton,” Virtual Reality Short Papers and Posters (VRW), pp. 119-120, 2012.

10.H. Y. Huang and S. H. Chang, “A Skeleton-Occluded Repair Method from Kinect,” International Symposium on Computer, Consumer and Control (IS3C), pp. 264-267, 2014.