運用3D運動特徵於人體姿態辨識及其學習方法之研究

A Study on 3D Motion Feature-Based Gesture Recognition

張哲維(Che-Wei Chang)

指導教授 ： 丁英智

國立虎尾科技大學/電機資訊學院/電機工程研究所/碩士(2014年)

摘要

本論文使用微軟公司的KINECT裝置開發出一套人體姿態辨識系統，並利用微軟提供之應用程式介面(Application Programming Interface, API)與軟體發展工具(Software Development Kit, SDK)擷取人體骨架座標及其3D運動特徵。在3D運動特徵的特徵值擷取部份，本論文以人體骨架座標為基礎，發展具關節點座標與關節間角度特性的特徵擷取方法，做為人體姿態辨識演算法之特徵值。本論文運用三種方法進行人體姿態辨識，分別是動態時間校正方法(dynamic time warping, DTW)、隱藏式馬可夫模型(hidden Markov model, HMM)與主成份分析法(Principal components analysis, PCA)，並且，三種人體姿態辨識方法再各自發展辨識系統學習方法。
在DTW人體姿態辨識方式，本論文提出內插學習之監督式學習方法，此方法可藉由一監督者的判定以決策是否將學習資料導入至參考樣本資料庫(Referenced templates)學習，所發展的內插學習方法並可調整辨識系統的學習速度。姿態辨識實驗之實驗結果證實此融入學習機制的DTW方法可有效提升傳統DTW方法的辨識準確率。
針對在HMM人體姿態辨識，本論文運用HMM進行人體姿態辨識，建立HMM人體姿態辨識系統，在HMM人體姿態辨識中並進行機器學習方法的研究。在HMM人體姿態辨識的系統學習中，一種沒有學習資料的HMM姿態模型觀察有學習資料的鄰近模型的學習狀況而亦可進行模型學習的自主學習演算法被發展。在所發展的基於HMM的人體姿態辨識系統學習方法中，所有模型皆可參照鄰近模型之移動方向與移動量進行相對應的學習，實驗結果顯示此方法可有效原HMM辨識系統的辨識率。
PCA廣泛應用於圖形識別領域，本論文運用此方法於姿態辨識，並發展出機器學習方法。在PCA人體姿態辨識研究中，一個完整動作的姿態樣本即可在特徵空間中找到其位置。所搜集的各式動作的姿態樣本資料庫可供人體姿態之3D運動特徵的特徵空間建立。由於使用者不同而有不同的動作習慣，其所造成之讓資料庫中的每一個姿態樣本長度不盡相同的問題可以在本論文中所設計的一種姿態樣本長度補齊方法中被克服。該方法將所有姿態進行長度統一化，因而可以透過PCA建立出人體姿態之3D運動特徵的特徵空間。本論文進行PCA人體姿態辨識之系統學習方法的研究，其主要為將學習資料加入原有資料庫中，並重建出新的特徵空間。實驗結果證實PCA人體姿態辨識與其學習方法具備良好的辨識性能。

關鍵字

KINECT ； 人體骨架 ； 姿態辨識 ； 動態時間校正 ； HMM ； PCA ； 機器學習

並列摘要

This thesis used Microsoft's KINECT device to develop a system of human gesture recognition. With the utilization of application programming interface (API) and software development kit (SDK) provided by Microsoft, extractions of the coordinates of the human skeleton and 3D-(x, y, z) motion feature are easy, which facilitates the recognition task. In the extraction work of 3D motion features, this thesis developed joint coordinate and joint angle approaches based on the Kinect-derived human skeletons for gesture recognition. This thesis uses three gesture recognition methods, dynamic time warping (DTW), hidden Markov model (HMM) and principal components analysis (PCA), and recognition system learning among all of these three recognition methods is further developed.
In DTW human gesture recognition, this thesis presents interpolation learning, which belongs the class of supervised learning. DTW interpolation learning can be judged by the supervisor's decision to decide whether the learning data is used for DTW reference template revisions. The proposed interpolation learning method can adjust learning speed of recognition systems. Experimental result shows the superiority of the proposed DTW learning method.
For HMM human gesture recognition, the thesis builds HMM of human gesture recognition, and developed its learning method. In the learning scheme of HMM gesture recognition, HMM gesture model without any learning data can be adjusted by observing the learning condition of the close HMM models. In HMM gesture recognition with system learning, all HMM models can be adjusted regardless of the learning data. Experiment result shows the effectiveness of the proposed HMM-based gesture recognition approach.
For PCA human gesture recognition, an active gesture sample performed by a human actor can find its position in the eigenpace. The eigenpace for describing the human active gestures is established through the collect gesture samples. Since the active habit of the user is different, the time duration of each active gesture is therefore different. This problem can be conquered by the data alignment method developed in this thesis. For PCA gesture recognition, recognition system learning is also investigated for further improving the recognition performance of PCA gesture recognition systems. Experiment result shows the effectiveness of PCA gesture recognition with system learning.

並列關鍵字

KINECT ； Human skeleton ； Gesture recognition ； DTW ； HMM ； PCA ； Machine learning

參考文獻

1. C. M. Tseng, C. L. Lai, D. Erdenetsogt and Y. F. Chen, “A microsoft kinect based virtual rehabilitation system,” International Symposium on Computer, Consumer and Control, pp. 934-937, June, 2014.

2. T. Y. Lin, C. H. Hsieh and J. D. Lee, “A kinect-based system for physical rehabilitation: utilizing tai chi exercises to improve movement disorders in patients with balance ability,” Proc. Asia Modelling Symposium, pp. 149-153, July, 2013.

3. A. K. Roy, Y. Soni and S. Dubey, “Enhancing effectiveness of motor rehabilitation using kinect motion sensing technology,” Proc. Global Humanitarian Technology Conference: South Asia Satellite, pp. 298-304, August, 2013.

4. C. Lee, H. Park, J. Kim, B. Kim, L. Kim and G. H. Kwon, “Motor rehabilitation based on brain machine interface and microsoft kinect,” Proc. IEEE International Conference on Consumer Electronics, pp.236-239, Jan, 2014.

5. A. Corradini, and H. Gross, “Camera-based gesture recognition for robot control,” Proc. IEEE-INNS-ENNS International, vol. 4, pp. 133-138, 2000.