使用Kinect體感攝影機藉由人體骨架進行人類動作識別

Using Human Skeleton to Recognizing Human Motion by Kinect's Camera

吳貴崗(Kuei-Kang Wu)

指導教授 ： 張厥煒

國立臺北科技大學/電資學院/資訊工程系研究所/碩士(2011年)

https://doi.org/10.6841/NTUT.2011.00007

摘要

人類動作的分析和識別是現行研究電腦視覺及圖形識別等領域的重要議題之一。人類的動作是由連續靜態姿勢構成，因為動作本身在空間與時間上具有相當複雜的高維度資訊，且動作可能會產生自我遮蔽(Self-Occlusion)現象，因此在一般的2D攝影機要精確地分析動作仍然有其瓶頸。在表達人體全身這方面，直覺上來說，以人體骨架來代表比用人體外型輪廓更為直觀。因此在本論文，使用Kinect體感攝影機進行人體追蹤及動作捕捉，進而找出人體骨架的關節三維直角座標資訊來進行人類動作識別。
本論文建立一套以視覺角度不變(View-Invariant)的動作比對模式。本論文提出使用Rule-Based混合Example-Based的方法。使用Kinect搭配OpenNI所產生的人體骨架15個關節的三維直角座標( x, y, z )值，座標經過正規化後，得到一組以身體軀幹為原點的物體座標系統。本論文以棒球運動為例，使用人體骨架關節之間的關係來估計姿勢，建立投手高壓投球規則，作為樣本比對前的濾器(Filter)，濾除不可能為投手高壓投球的動作，再使用現場動作跟從訓練樣本中以K-Means分群演算法算出的最佳5個關鍵姿勢樣本比對15個關節的三維直角座標，計算現場動作跟最佳5個關鍵姿勢樣本的15個關節的歐幾里得距離，其總和即為動作差距，動作差距越小則相似度越高。
　　此一動作比對模式無論在辨識的平均相似度為86.10%及執行速度為25 FPS，都有不錯的效果，可推廣至其他動作的識別。

關鍵字

人體骨架 ； 人類動作 ； 動作識別 ； Kinect ； Rule-Based ； Example-Based

並列摘要

This paper establishes a motion matching model of View-Invariant. This paper presents the use of Rule-Based and Example-Based approach. Using Kinect withOpenNI to generated three-dimensional Cartesian coordinates (x, y, z) value of the human skeleton joints, and then using the translation method of the geometrytransfer to normalized coordinates. After normalized coordinates, we get the objects coordinate system in which used the torso center as the origin.
This paper used baseball as example. In order to creating the rules of overhand pitching. We used the relationship between the joints position of human skeleton to estimating motion postures. Before motion matching, used these rules as the filter to filtering those posture which can’t be overhand pitching. Then use the live motion to matching the 3D coordinates of 15 joints of the best sample which are calculated and selected from training samples by K-Means clustering algorithms. Calculating the3D coordinates distances between 15 joints of live motion skeleton and the bestsample skeleton. The smaller distance the higher similarity.
This paper establishes the motion matching model of View-Invariant has good results in terms of average recognize similarity is 86.10% and process speed is 25FPS. It can be extended to recognize the other athletics motion.

並列關鍵字

Human Skeleton ； Human Motion ； Motion Recognition ； Kinect ； Rule-Based ； Example-Based

參考文獻

[9] Xiaofei Ji, Honghai Liu, "Advances in View-Invariant Human Motion Analysis: A Review", IEEE Transactions On Systems, Man, And Cybernetics—Part C: Applications And Reviews, Vol. 40, No. 1, January 2010

[10] Daniel Weinland, Remi Ronfard, Edmond Boyer, "A Survey of Vision-Based Methods for Action Representation, Segmentation and Recognition", INRIA, Feb. 2010

[11] Ronald Poppe, "Vision-based human motion analysis: An overview", Computer Vision and Image Understanding 108, 2007, pp.4–18

[13] Y.C.Wu, H.S.Chen, W.J.Tsai, S.Y.Lee, J.Y.Yu, "Human Action Recognition Based On Layered-HMM", 2008 IEEE international conference on multimedia and expo ICME 2008 proceedings

[14] Aaron F. Bobick, James W. Davis, "The Recognition of Human Movement Using Temporal Templates", IEEE Transactions On Pattern Analysis and Machine Intelligence, Vol. 23, No. 3, March 2001