在本篇論文中，我們提出了能夠偵測遮蔽行人的方法，將行人分成頭部、身體、腿部三個部位，每個部位的姿勢皆不相同，根據各個部位所能運動的自由度，等量量化後產生不同的姿勢，頭部分成中間、右邊、左邊三種姿勢，身體分成正身和側身兩種姿勢，腿部則是由左右腳的彎曲程度分成八種姿勢，將定義的部位和各個姿勢，分別訓練專屬的偵測器。
在特徵擷取部分利用梯度方向直方圖(HOG)特徵，結合串聯式AdaBoost的分類演算法來訓練分類器，梯度方向直方圖(HOG)能夠成功的擷取行人的輪廓和邊緣資訊。AdaBoost是以特徵選取為基礎的分類演算法，從一大群弱分類器(weak classifier)，找出具有鑑別性的特徵，並給予弱分類器各自的權重，成為一個強分類器(strong classifier)。串聯(cascade)的結構能使得分類更有效率，將一個大分類器分成多層的小分類器串聯，大部分屬於非行人的影像能夠在前幾層就被排除，能夠通過最後一層才會被判斷為行人。
在遮蔽處理部分，將行人各個部位的偵測分數，給予不同的權重重新組合後，建立遮蔽配置圖(occlusion map)，從中找出最可能發生的遮蔽情形，被遮蔽的部位偵測器不列入考量中，其餘可見部位再由剩餘的偵測器分數平均，分數最高的遮蔽情形即為最佳配置，對每個行人重新給分數，透過以上的方法，就可以判斷出行人或是非行人的影像，此外，我們使用了兩個資料庫INRIA dataset[16]和Caltech dataset[17]來訓練與測試實驗數據。

In the thesis, we propose a method to detect partially occluded pedestrians. The human body can be represented by three part regions—head, torso, and pair of upper legs. Due to highly articulated human poses and varying viewing angles, we segment them according to degree of freedom of human pose space. The head is divided into three positions — middle, right, and left. The torso is divided into two groups — front/back view, and side view. The leg is divided into eight kinds of poses. The human part detector trained using above poses. In feature extraction and classification, we use Histograms of Oriented Gradients (HOG) feature combined with AdaBoost cascade algorithm to train the classifier. HOG feature can successfully capture the contour and edge information. Cascade structure can make the classification more efficient.
Part-based detectors have demonstrated their merit in partially occluded human detection. However, there is a key issue to be solved on how to integrate the scores of part detectors. We build occlusion map to find the most likely occlusion type. The highest merging score is the best configuration and reevaluate the detection score of each human. Experimental results on two public datasets (INRIA and Caltech) show the effectiveness of the proposed approach.

Chapter 1 Introduction 1
1.1 Motivation 1
1.2 Related works 2
1.3 System overview 3
1.4 The organization of this thesis 4
Chapter 2 Learning the human pose 5
2.1 Body part segmentation 5
2.2 Learning the part pose 6
2.3 Pose with occlusion 7
Chapter 3 Adaboost cascade algorithm 9
3.1 The boosting algorithm 9
3.2 Cascade of classifiers 10
3.3 The training algorithm for building a cascaded detector 11
Chapter 4 Pose-adaptive detector 13
4.1 HOG feature extraction 13
4.2 Variable-size blocks 14
4.3 Non-maximum suppression 15
4.4 Mean shift mode detection algorithm 16
Chapter 5 Part detector for occlusion handling 19
5.1 Part combination detector 19
5.2 Occlusion handling 20
Chapter 6 Experimental results 22
6.1 Training and testing dataset 22
6.2 Evaluation methodology 23
6.3 Cascade details 24
6.4 Performance evaluation 27
6.5 Comparison of evaluated pedestrian detector 33
Chapter 7 Conclusion and future works 35
References 36

[1] N. Dalal, B. Triggs. “Histograms of oriented gradients for human detection,” IEEE Conf. Computer Vision and Pattern Recognition, vol. 1, pp. 886-893, July 2005.
[2] Xiaoyu Wang, Tony X. Han, Shuicheng Yan. “An HOG-LBP human detector with Partial occlusion handling,” IEEE Conf. Computer Vision, pp.32-39, 2009.
[3] P. Felzenszwalb, D. McAllester, and D. Ramanan, “A Discriminatively Trained, Multiscale, Deformable Part Model,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2008.
[4] Z. Lin and L.S. Davis, “A Pose-Invariant Descriptor for Human Detection and Segmentation,” Proc. European Conf. Computer Vision, 2008.
[5] S. Maji, A. Berg, and J. Malik, “Classification Using Intersection Kernel Svms Is Efficient,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2008.
[6] W. Schwartz, A. Kembhavi, D. Harwood, and L. Davis, “Human Detection Using Partial Least Squares Analysis,” Proc. IEEE Int’l Conf. Computer Vision, 2009.
[7] P. Dollar, Z. Tu, P. Perona, and S. Belongie, “Integral Channel Features,” Proc. British Machine Vision Conf., 2009.
[8] S. Walk, N. Majer, K. Schindler, and B. Schiele, “New Features and Insights for Pedestrian Detection,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2010.
[9] P. Dollar, S. Belongie, and P. Perona, “The Fastest Pedestrian Detector in the West,” Proc. British Machine Vision Conf., 2010.
[10] C. Wojek and B. Schiele, “A Performance Evaluation of Single and Multi-Feature People Detection,” Proc. DAGM Symp. Pattern Recognition, 2008.
[11] P. Sabzmeydani and G. Mori, “Detecting Pedestrians by Learning Shapelet Features,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2007.
[12] A. Bar-Hillel, D. Levi, E. Krupka, and C. Goldberg, “Part-Based Feature Synthesis for Human Detection,” Proc. European Conf. Computer Vision, 2010.
[13] P. Dollar, Z. Tu, H. Tao, and S. Belongie, “Feature Mining for Image Classification,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2007.
[14] P. Dollar, C. Wojek, B. Schiele, and P. Perona. “Pedestrian detection: An evaluation of the state of the art,” TPAMI, 2011.
[15] P. Viola and M. Jones. Robust real-time face detection,” In IJCV, 2004.
[16] Y. Freund and R. E. Schapire. “A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting,” Journal of Computer and System Sciences, 1997.
[17] INRIA Person Dataset, http://pascal.inrialpes.fr/data/human/
[18] Caltech Pedestrian Dataset, http://www.vision.caltech.edu/Image_Datasets/