臺灣博碩士論文加值系統

高斯混合模型和隱藏式馬可夫模型常用來估計資料的隱藏狀態。高斯模型假設資料間的獨立性而馬可夫模型容許時間序列資料間存在馬可夫的關係。但是馬可夫模型假設下隱藏狀態的估計需要有效的演算，譬如 EM 演算法中要求高收斂速度，也希望結果不致於落入區域而非全域的極值。對於影片資料中前景和背景的分離，我們可以看作是一個隱藏狀態的估計問題。向來研究多利用兩個單一模型分別來估計此一隱藏狀態，即一影片位置是否為前景或背景。我們提出一個綜合高斯模型和隱藏式馬可夫模型兩單一模型優點的整合生成模型來分離動態背景中的前景，並透過一些實驗進一步了解這個整合生成模型的優勢所在。實驗的結果說明，當採用整合模型的方法時，EM 演算法中收斂緩慢的情形較少。相對的，如果使用隨機的初始值來訓練隱藏式馬可夫模型，那麼將會產生許多收斂緩慢的情形。另外，在 EM 演算法平均迭代的次數上，當採用整合模型的方法時，平均迭代的次數相對的比較少。另外在效能的提升上，我們使用像素取樣的方法來訓練我們的模型。此法不但可以提升效率而且可達到一定的正確率，增加整體的可用性。除此之外我們也將空間上相鄰像素一起考慮來改良模型的參數估計時，這樣做不但可以得到像素取樣方法的優點，也大幅的提升了整體的正確率。

As we know, the Gaussian Mixture Model and the Hidden Markov Model are usually used for the estimation of hidden states. Gaussian Mixture Model assumes the independence between data while Hidden Markov Model allows Markov assumption in a data series. But Hidden Markov Model needs effective computation where for instance, we would like EM algorithm to converge rapidly and better to converge to the global instead of local optimum. To segment the moving object in front of non-stationary background in a movie, many researchers apply either Mixture Model or Hidden Makorv Model to estimate the hidden state, i. e. the location to be an object or background. We propose an Integrated Generative Model which combines two previous models and try to understand the advantage of this model through various empirical studies. When we use the Integrated Model, we find the slow converging cases in the EM algorithm are much lower than, using the random initial parameters to train the Hidden Markov Model. Besides, the converging speed is enhanced, when we use the Integrated Model. To improve our model, we further apply a sampling approach in our training. By that, not only the training is faster than before and the accuracy increases. We also consider the neighboring pixels together on the spatial domain to improve training of the model. Combined with the sampling skill, adding neighboring pixels can raise the accuracy and increase the efficiency at the same time.

第一章 緒論 1
1.1 前言 1
1.2 文獻回顧 1
1.3 方法論 2
1.3.1 高斯混合模型 (Gaussian Mixture Model) 2
1.3.2 模型選擇 (Model Selection) 4
1.3.3 隱藏式馬可夫模型 (Hidden Markov Model) 5
1.3.4 整合生成模型 (Integrated Generative Model) 9
1.4 論文大綱 12
第二章 整合生成模型 (Integrated Generative Model) 13
2.1 基本概述 15
2.2 整合生成模型的分析 17
2.2.1 比較不收斂位置的百分比 17
2.2.2 比較平均迭代的次數 22
2.2.3 在不同的背景範圍比較平均迭代的次數 23
2.2.4 結論 32
2.3 比較兩種整合生成模型之間的狀態估計 33
2.3.1 在訓練資料組的比較 34
2.3.2 在測試資料組的比較 37
2.3.3 結論 40
第三章 效能提升與改良 41
3.1 像素取樣 (Sampling) 41
3.1.1 在訓練資料組的比較 43
3.1.2 在測試資料組的比較 45
3.1.3 結論 48
3.2 調整視窗大小 (Windows Size) 49
3.2.1 在訓練資料組的比較 50
3.2.2 在測試資料組的比較 52
3.2.3 結論 54
第四章 結論 55
4.1 未來工作 56
參考文獻 58

[1] G. Welch and G. Bishop, An Introduction to the Kalman Filter, an introduction, Department of Computer Science, University of North Carolina, 2006
[2] C. Stauffer and W.E.L. Grimson, Adaptive Background Mixture Models for Real-Time Tracking, Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'99), vol.2, pages 246-252, 1999
[3] G. Rigoll, S. Eickeler, and S. Muller, Person Tracking in Real-World Scenarios Using Statistical Methods, Fourth IEEE International Conference on Automatic Face and Gesture Recognition, pages 342-347, 2000
[4] J. Rittscher, J. Kato, S. Joga, and A. Blake, A probabilistic background model for tracking, Proceedings of the 6th European Conference on Computer Vision, pages 336-350, 2000
[5] P. KaewTraKulPong and R. Bowden, An Improved Adaptive Background Mixture Model for Realtime Tracking with Shadow Detection, Proceedings of the 2nd European Workshop on Advanced Video Based Surveillance Systems, 2001
[6] J. Zhong and S. Sclaroff, Segmenting foreground objects from a dynamic textured background via a robust Kalman filter, Proceedings of the Ninth IEEE International Conference on Computer Vision, vol. 1, pages 44-50, 2003
[7] C. Tomasi, Estimating gaussian mixture densities with em, a tutorial, 2003
[8] Y.H. Lee, Objects Tracking in Non-stationary Environment, a master thesis, Department of Computer Science and Information Engineering, National Taiwan University of Science and Technology, 2006
[9] A. Krogh, G. Mitchison, R. Durbin, and S. Eddy, Biological sequence analysis, The press syndicate of the university of Cambridge, 1998
[10] T. Xiang and S. Gong, Beyond tracking: Modeling activity and understanding behaviour. International Journal of Computer Vision, 67(1):21–51, 2006