結合高度與形狀資訊於多視角動作辨識

Multiple Viewpoints Action Recognition Based on Height and Shape Information

孫培強(Pei-Chiang Sun)

指導教授 ： 陳永耀

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.01475

摘要

近年來，由於人們對於人機互動的渴求，愈來愈多研究投入於人類動作辨識系統的研發。然而，為了能增加動作辨識系統的適用性，該系統必須能夠適用在不同的視角下。因此，本論文提出了一個能夠適用於多種視角的動作辨識系統。
為了達成動作辨識，在此利用了姿態的序列進而推論動作。一般而言，人類可以藉由觀察人體的姿態推論可能發生的動作。基於此概念，姿態與各種動作間的相關性可以量化表示成一個向量，稱之為權重向量。本論文所提出的系統便是基於權重向量的概念進而判斷動作。
然而，因為單一影像攝影機所擷取的形狀特徵會受到視角影響，系統無法僅利用人體的形狀資訊在不同視角下正確地辨識人體姿態；相反地，人體的高度分布卻不會因為視角的變化而有所改變。換言之，人體的高度分布可以作為系統辨識姿態的另一個有利特徵。因此，採用了能夠同時拍攝影像和深度資訊的攝影機以擷取出人體的形狀與高度資訊，使系統正確地辨識人體姿態，進而提高動作辨識率。
本論文在考慮視角變化下用16種不同動作進行系統測試，最後得到了96.9%的平均辨識率。

關鍵字

動作辨識 ； 多視角 ； 關鍵姿態 ； 人體高度分布 ； 深度

並列摘要

Researches for human action recognition system become more and more popular in recent years since people want to make machine more interactive and friendly. To increase the applicability of the action recognition system, the system should recognize human actions under view changes. Thus, this thesis proposes a view-invariant action recognition system which can recognize human actions under view changes.
For the action recognition mechanism, the proposed system uses a sequence of postures to infer human actions. Usually, human can perceive actions by observing only the human body postures. Inspired by this property, the relativity between a posture and various actions can be represented as a vector which is called weighting vector. Then, the propose system can infer human actions based on the weighting vectors of postures.
However, because of the viewing angle effect of a single camera, the system cannot recognize the human posture correctly under view changes by only using shape information; inversely, the height distribution of human body would not vary under view changes. Namely, the height distribution of human body is another significant feature for posture recognition. Thus, in this thesis, a RGB-Depth camera is used to extract the shape and the height information from human body for posture recognition to improve the performance of the proposed system.
The proposed system has been tested by sixteen kinds of actions under view changes and achieve 96.9% recognition rate.

並列關鍵字

Action recognition ； multiple viewpoints ； key posture ； height distribution of human body ； depth map

參考文獻

[2] S.-T. Su, "Moving Object Detection Based on Two-Staged Background Subtraction Approach," Master, Department of Electrical Engineering College of Electrical Engineering and Computer Science, National Taiwan University, 2009.

[3] Y. Wang and G. Mori, "Hidden Part Models for Human Action Recognition: Probabilistic versus Max Margin," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, pp. 1310-1323, 2011.

[4] I. N. Junejo, E. Dexter, I. Laptev, and P. Perez, "View-Independent Action Recognition from Temporal Self-Similarities," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, pp. 172-185, 2011.

[5] P. Dollar, V. Rabaud, G. Cottrell, and S. Belongie, "Behavior recognition via sparse spatio-temporal features," presented at the 2nd Joint IEEE International Workshop on Visual Surveillance and Performance Evaluation of Tracking and Surveillance, 2005.

[6] S. Ali and M. Shah, "Human Action Recognition in Videos Using Kinematic Features and Multiple Instance Learning," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 32, pp. 288-303, 2010.