近年來，由於電腦硬體設備、電腦視覺與機器學習技術的演進，使得行為辨識與影像辨識的技術不斷進步，且被廣泛應用在現實生活中，其中，人類行為辨識與人臉辨識為電腦視覺結合機器學習的應用中相當廣泛的一項議題。

本論文主要是用人臉物件偵測、人臉物件追蹤、影像特徵值擷取與機器學習技術開發出一套全自動即時電腦視覺結合多階層式人工智慧系統來應用於臉部美妝保養行為辨識，由於本系統擁有多類別性、影片裡的影像出現頻率不均、資料變異性高。這些問題在行為辨識的技術開發上具備相當高的挑戰性。

本研究利用自身建立的影片資料庫，此影片資料庫擁有多動作類別影片與動作類別資料量分布不均的特性，對此影片資料庫裡每個動作影像序列進行人臉物件偵測與追蹤來擷取出人臉區域視窗，之後調整人臉區域視窗到合適的大小來產生區域影像，對所有區域影像做影像特徵值擷取演算法得到三維MHI影像特徵值，然後利用三維MHI影像特徵值資料庫來結合機器學習演算法隨機森林(random forests)和多階層式AdaBoostM1來做演算法分析，在演算法分析的結果中發現多階層式AdaBoostM1演算架構能有效與精確地辨識行為。

In recent years, due to rapid development of computer instruments
and computer vision and machine learning techniques, the action
recognition technology has been progressed constantly. Action
recognition techniques can be applied widely into the computer
vision applications with artificial intelligence in the human
living environment. Two of the most popular research topics in the
field of computer vision with machine intelligence are human
action recognition and face recognition.

In this thesis, a real time computer vision action recognition
system with hierarchical machine learning models is presented. The
system consists of face detection, face tracking, feature
extraction and action recognition techniques and can be applied to
recognizing facial makeups behaviors. It is extremely challenging
to develop this system because of the complexity of the targeting
applications, which is constituted of 77 kinds of actions for
recognition. Apart from high dimensionality of the data to deal
with, real time computation is critical for live streaming data
processing.

In this work, a real time action recognition system has been
built. A large video dataset were constructed for training and
quantitative evaluation. The proposed system conducts face
detection and tracking to locate the facial region of interests
for each video frame. Then, the system normalizes the size of each
region of interests and applies feature extraction to extract 3D
Motion History Image (MHI) features. In experiments, we compare
two different machine learning algorithms and learning frameworks,
including a single layer random forest and a multi-layer
hierarchical AdaBoostM1 ensemble model. In evaluation, we found
that the multi-layer hierarchical AdaBoostM1 model performs more
efficiently and effectively, producing the models only one-fifth
of the size that the single layer random forest model requires.

[1] Y. Lu, Y. Wang X. Tong, Z. Zhao, H. Jia and J. Kong. Face Tracking in
Video Sequences Using Particle Filter Based on Skin Color Model and Facial
Contour Second International Symposium on Intelligent Information Technology
Application vol.1, pp. 451-461 (2008).
[2] R. C. Verma, C. Schmid and K. Mikolajczyk Face detection and tracking in
a video by propagating detection probabilities IEEE Transactions on Pattern
Analysis and Machine Intelligence vol.25, no.10, pp. 1215-1228 (2003).
[3] Raul Humberto Pena-Gonzalez, Marco Aurelio Nuno-Maganda Computer vision
based real-time vehicle tracking and classi cation system. 57th IEEE Interna-
tional Midwest Symposium on Circuits and Systems , pp. 679-682 (2014).
[4] E. R. Davies Computer and Machine Vision Academic Press, New York , pp.
679-682 (2012).
[5] J. Canny A computational approach to edge detection IEEE Trans. Pattern
Analysis and Machine Intelligence , pp. 679-698 (1986).
[6] O.J. Tobias and R. Seara Image Segmentation by Histogram thresholding using
fuzzy sets IEEE Trans. Date of Publication , pp. 1457-1465 (2002).
[7] R. Brunelli and T. Poggio Face recognition: features versus templates IEEE
Trans. Pattern Analysis and Machine Intelligence , pp. 1042-1052 (1993).
[8] S. Ali and M. Shah Human action recognition in videos using kinematic features
and multiple instance learning IEEE Trans. Pattern Analysis and Machine In-
telligence Vol.32(2010).
[9] T. Zhang and S. Liu Boosted multi-class semi-supervised learning for human
action recognition Pattern Recognition Vol.44,pp.23334-2342(2011).
[10] J. Greenhalgh and M. Mirmehdi Real-Time Detection and Recognition of Road
Tra c Signs IEEE Transactions on Intelligent Transportation Systems vol.13,
no.4, December(2012).
[11] J. Matas Robust wide-baseline stereo from maximally stable extremal regions
Image Vis. Comput. vol.22, no.10, pp. 761-767, Sep.(2004).
[12] N. Dalal and B. Triggs Histograms of oriented gradients for human detection
Proc. CVPR , pp. 886-893(2005).
[13] C. Cortes and V. Vapnik Support vector networks J. Mach. Learn. vol.20,
no.3, pp. 273-297, Sep.(1995).
[14] A. F. Bobick and J. W. Davis The recognition of human movement using tempo-
ral templates IEEE Transactions on Pattern Analysis and Machine Intelligence
vol.23, no.3, pp. 257-267 (Mar. 2001).
[15] A. F. Bobick and J. W. Davis An Appearance-Based Representation of Action
Proc. Int'l Conf. Pattern Recognition , pp. 307-312 (1996).
[16] M. Hu Visual Pattern Recognition by Moment Invariants IRE Trans. Infor-
mation Theory vol.8, no.2, pp. 179-187 (1962).
[17] Juergen Gall, A. Yao, N. Razavi, L. V. Goal and V. Lempitsky Hough Forests
for Object Detection, Tracking, and Action Recognition IEEE Transactions on
Pattern Analysis and Machine Intellignece vol.33, No.11, November(2011).
[18] P. Dollar, V. Rabaud, G. Cotrell, and S. Belongie Behavior recognition via
sparse spatio-temporal features VS-PETS (2005).
[19] M. Isard and A. Blake Contour tracking by stochastic propagation of condi-
tional density European Conf. Computer Vision , pp.343-356(1996).
[20] A. Doucet, N. De Freitas, and N. Gordon Sequential Monte Carlo Methods in
Practice New York: Springer (2001).
[21] C.W. Wang, A. Hunter, N. Gravill,and S. Matusiewicz Unconstrained Video
Monitoring of Breathing Behavior and Application to Diagnosis of Sleep Apnea
IEEE Transactions on Biomedical Engineering vol.61, No.2, February(2014).
[22] A. Makarov Comparison of background extraction based intrusion detection
algorithms Proc. Int. Conf. Image Process. vol.1,pp. 521-524,February(1996).
[23] N. T. Sibel and S. J. Maybank Fusion of multiple tracking algorithms for robust
people tracking Proc. Eur. Conf. Comput. Vis. vol.4,pp. 373-387(2002).
[24] P. Viola and M. Jones Rapid Object Detection using a Boosted Cascade of
Simple Features Conf. Computer Vision and Pattern Recognition ,(2001).
[25] C. Papageorgiou, M. Oren, and T. Poggio A general framework for object
detection International Conf. Computer Vision ,(1998).
[26] Y. Freund and R. E. Schapire A decision-theoretic generalization of on-line
learning and an application to boosting Computational Learning Theory: Euro-
colt '95 Springer-Verlag,pages 23-27(1995).
[27] R. E. Schapire, Y. Freund, P. Bartlett and W.S. Lee Boosting the margin:
A new explanation for the e ectiveness of voting methods International Conf.
Machine Learning (1997).
[28] G.R. Bradski Real time face and object tracking as a component of a per-
ceptual user interface Application of Computer Vision,WACV'98,Proceedings,
IEEE Workshop on ,pp. 214-219,Oct(1998).
[29] J. Y. Bouguet Pyramidal Implementation of the Lucas Kanade Feature Tracker
Description of the algorithm Microprocessor Research Labs,Intel Corporation
(2000).
[30] Y. Freund and R. E. Schapire Experiments with a New Boosting Algorithm
13th International Conf. Machine Learning (1996).
[31] J. R. Quinlan C4.5 : Programs for machine learning (1993).
[32] Weka http://www.cs.waikato.ac.nz/ml/weka/ .
[33] L. Breiman Random Forests Machine Learning no.45,5-32(2001).
[34] Dietterich, T. An experimental comparison of three methods for constructing
ensembles of decision trees: Bagging, boosting and randomization Machine
Learning 1998 1-21
71