應用Kinect之人體多姿態辨識

Human''s posture recognition by using Kinect

黃識夫(Shih-fu Huang)

指導教授 ： 王文俊

國立中央大學/資訊電機學院/電機工程學系/碩士(2011年)

摘要

本論文的研究目的為應用裝設於房間內的 Kinect 感測器擷取目標人體，辨識五種人體姿態，並應用水平投影量、星狀骨骼化、類神經網路、相似特徵處理的技術，來加以辨識五種人體姿態，站姿、坐姿、彎腰、跪姿以及躺姿。當Kinect 抓取到人體影像後，藉由深度資訊的改變，把人體形狀輪廓從背景中分離出來。接著使用水平投影量來辨識姿態是否為跪姿；如不為跪姿，則使用星狀骨骼化計算出五組由人體重心點至輪廓特徵點的特徵向量。五組的特徵向量與深度資訊為 Learning Vector Quantization (LVQ) [20]類神經網路的輸入，以訓練姿態辨識的權重值；接著LVQ網路的輸出可辨識站姿、正向坐姿、非正向坐姿、彎腰以及躺姿。對於站姿與非正向坐姿，將會經過相似特徵處理程序，藉由水平與垂直投影量，濾除手部干擾，計算出人體長寬比再加以辨識，以提高辨識率。姿態辨識系統在不同的室內環境下，對於不同體型、距離 Kinect 遠近不同的人體，都可達到即時且穩定的姿態辨識。因此辨識系統可實際應用在居家看護與遊憩環境照顧。

關鍵字

人體姿態辨識 ； Kinect

並列摘要

The objective of this study is to recognize human’s five postures which are captured from the set of Kinect. By using horizontal projection, star skeleton, neural network, and similar feature process techniques, five human’s postures, which contain standing, sitting, bending, keeling and lying, are recognized. After Kinect captures the picture of a human, a silhouette contour of the human is segmented from the background based on the difference of depth data between the human’s body and background. Then the horizontal projection is utilized to identify whether the posture is keeling or not. If it is not a kneeling posture, a star skeleton is used to calculate five maximum distances from the feature points to the centroid of the human body. The five branches of the star skeleton and depth data are the inputs to train the network of Learning Vector Quantization (LVQ). Subsequently, the outputs of the LVQ are utilized to recognize the five postures including standing, forward sitting, non-forward sitting, bending, and lying. The standing and non-forward sitting postures are processed by the similar feature process based on the horizontal and vertical projection. The hand-shaking disturbance is filtered to calculate the length and breadth ratio of human so as to improve the ratio of posture recognition. The posture recognition system can not only be applied to different indoor environments and different distances between Kinect and human, but also achieve the goal of real-time and stable posture recognition for different human physiques. Therefore, the system can be practically applied to home nursing and amusement place care.

並列關鍵字

Posture recognition ； Kinect

參考文獻

[3] 李乾丞(陳永耀教授指導), ”Fast Human Posture Recognition by Heuristic Rules”, 國立台灣大學碩士論文，2006年。

[4] C. C. Li, and Y. Y. Chen, “Human posture recognition by simple rules,” in Proceedings of IEEE Systems Man Cybernetics, Oct. 2006, pp. 3237-3240.

[5] B. Castiello, T. D’Orazio, A. M. Fanelli, P. Spagnolo, and M. A. Torsello, “A model-free approach for posture classification,” in Proceedings of IEEE Advanced Video and Signal Based Surveillance, Sep. 2005, pp. 276-281.

[7] J. W. Hsieh, C. H. Chuang, S. Y. Chen, C. C. Chen, and K. C. Fan, “Segmentation of human body parts using deformable triangulation,” IEEE Transaction on System, Man, and Cybernetics, Part A: Systems and Humans, vol. 40, no. 3, pp. 596-610, May. 2010.

[8] C. H. Chuang, J. W. Hsieh, L.-W. Tsai, and K. C. Fan, “Human action recognition using star templates and Delaunay triangulation,” in Proceedings of IEEE Intelligent Information Hiding and Multimedia Signal Processing, 2008, pp. 179-182.