在雜亂與移動背景狀況下之視訊中的人體動作行為辨識

Human Action Recognition in Video underClutter and Moving Background

甘政中(Cheng-Chung Kan)

指導教授 ： 陳淑媛

元智大學/資訊學院/資訊工程學系/碩士(2013年)

https://doi.org/10.6838/YZU.2013.00152

摘要

針對視訊中的動作行為的識別一直是電腦視覺領域中的一個極為熱門且值得探討的一個問題。尤其是針對人類動作識別部分，而其應用範圍極為廣泛，包括視頻檢索，智慧型運輸系統控制，視訊監控、視訊中的事件分析和人機互動，促使視訊中的人類之行為識別成為電腦視覺領域中的一個熱門研究主題。但是目前大多數的人體行為研究所使用的視訊資料大都是基於固定的場景和簡單環境所拍得的，其動作結果較單純，然而在真實的場景中的畫面中則常因為人體行為動作複雜、背景雜亂、光照亮度變化、人體大小變化、攝影機視角變化與、及攝影機移動等問題而產生及複雜的視訊。
此次的研究工作將針對在雜亂與移動背景狀況下之視訊中人體動作行為辨識進行研究，並將以YouTube的資料庫為基本的測試視訊，基本上將基於局部時空點計算得到之特徵進行人體行為識別，並以提升識別正確率為目的。比對方式則將基於NBMIM方式的局部時空特徵的人體行為識別方法。所以研究方式包含有特徵點與時空興趣點選取分析、改良計算歸類給某一動作行為的機率計算方式，資料庫建立時的時空特徵向量與動作類別的重要性及相關性分析以選取具分辨能力的時空特徵並加重投票或計算機率的權重，此外並進行攝影機移動的向量之估計動作以進行時空特徵點的篩選，如此可以消除複雜動態背景之影響以取得穩定且重要的特徵點以達成更正確且具有韌性的人體動作行為識別方法。

關鍵字

人體動作識別 ； 局部時空特徵 ； 背景移動估算 ； 類別交互資訊 ； 類條件概率估計 ； 資料庫分類

並列摘要

Behavior on action recognition in video sequences has becoming a more interesting and active research for computer vision, especially for the for human action recognition. The applications and demands of the human action recognition are expanded extensively, such as video retrieval, intelligent transport system control, video surveillance and video event analysis. Generally the video data used by most of the researches are based on fixed camera position with stationary background scenes, well-controlled environment and simple actions. However, in real human action cases, the actions of human behavior are often complex and the background are filled with cluttered background with illumination brightness changes, different human body size, different camera view positions and angles and even with moving camera. These make the real video much more complex.
In this study, research issues are focused on improving accuracies of human action recognition under clutter and moving background conditions as mentioned above. The dataset used for testing is the YouTube dataset. The whole recognition scheme will be based on the well-known local space-time interest point (STIP) and the recognition scheme will be based on NBMIM. Research issues including (1) Reliable feature points selection (2) the improvement of the probability estimation for recognition classification (3) the analysis and correlations of the important STIP features in database training (4) the selection of stable and robust feature points based on camera motion estimation in order to achieve an more accurate and robust human action recognition method.

並列關鍵字

Human action recognition ； spatio-temporal features selection ； database clustering ； features correlation analysis ； camera motion estimation ； clutter moving background

參考文獻

[1] J. K. Aggarwal and M. S. Ryoo, “Human activity analysis: a review,” ACM Computing Surveys, vol. 43, no. 3, 2011.

[2] R. Roppe, “A survey on vision-based human action recognition,” Image and Vision Computing, vol. 28, no. 3, pp. 976–990, 2010.

[3] K. Hatun and P. Duygulu “Pose Sentences: A new representation for action recognition using sequence of pose words,” in Proc. Int. Conf. Pattern Recognition, 2008

[4] H. J. Seo and P. Milanfar, “Action recognition from one example,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 33, no. 5, pp. 867–882, 2011.

[5] N. Ikizler and P. Duygulu, “Histogram of oriented rectangles: A new pose descriptor for human action recognition,” Image and Vision Computing, vol. 27, no. 10, pp. 1515–1526, 2009