行人辨識主要目的在於跨鏡頭人物的搜尋，也就是在不同鏡頭中找到同一個人。在深度學習被廣泛應用前，使用人物特徵直接進行比對是種常見方法，本文中研究的顏色特徵，便是屬於其中的一種方式。將顏色代表的數值劃分，使用KNN方式歸類各個顏色所在區域，以此為依據，取得人物圖像的顏色特徵，進行辨別。但由於鏡頭解析度不同、光線明暗不同，使得該人物即使衣著不變的情況下，顏色的特徵也會產生變化，進而影響比對的結果。所以，本文另一研究，採用深度學習方式，先對人物圖像進行水平切割，分別提取特徵，進行特徵訓練，並且加入動態對齊的演算，用來彌補人物姿態不一致的問題。最終實驗結果，優於直接特徵擷取的比對準確度相對提高。

Person Re-Identification is mainly aimed at cross-camera to search the same person in different shots. The direct comparison is a common way which have been used to recognize feature before deep learning was applied widely. In this study, one of the methods to pedestrian detection is using the color feature. We use the KNN method to categorize the area where each color is located by dividing the values represented by the colors and use this as a basis to obtain the color features of the person's image for identification. However, due to the different resolution of the lens and the different light and shade, the color characteristics of the person will change even if the clothes remain unchanged, which will affect the comparison results. Therefore, in this paper adopts deep learning approach, firstly, figure images to cut the figure images horizontally, extract the features separately, and conduct feature training, and add dynamic alignment algorithm to level out the problem of inconsistent figure poses. The final experimental results show that the comparison accuracy is relatively improved, which is better than direct feature extraction.

摘要
ABSTRACT
誌謝
目錄
表目錄
圖目錄
壹 、序論
一、 研究動機
二、 論文架構
貳 、文獻探討
一、 基於顯著顏色名稱的顏色描述符
(一) 計算顏色描述
(二) 特徵表達
(三) 人物比對
二、 超越人類水平的行人辨識技術
(一) 網路結構
(二) 動態對齊
(三) 三元難損失
參 、訓練
一、 訓練資料集
二、 損失函數
三、 評估方法與訓練數據
肆 、實驗結果
一、 實驗環境
二、 實驗資料
三、 實驗結果
伍 、結論
參考文獻

[1]Y. Yang, J. Yang, J. Yan, S. Liao, D. Yi及S. Z. Li, 2014, “Salient Color Names for Person Re-identification“, 收入 Computer Vision – ECCV 2014, Cham, 2014, 頁 536–551.
[2]J. Wang, J. Yang, K. Yu, F. Lv, T. Huang及Y. Gong, 2010, “Locality-constrained Linear Coding for image classification“, 收入 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 6月 2010, 頁 3360–3367.
[3]Y. Huang, K. Huang, Y. Yu及T. Tan, 2011, “Salient coding for image classification“, 收入 CVPR 2011, 6月 2011, 頁 1753–1760.
[4]X. Zhang, Y. Yang, L. C. Jiao及F. Dong, 2013, “Manifold-constrained coding and sparse representation for human action recognition“, Pattern Recognit., 卷 46, 期 7, 頁 1819–1831, 7月 2013.
[5]L. Liu, L. Wang及X. Liu, 2011, “In defense of soft-assignment coding“, 收入 2011 International Conference on Computer Vision, 11月 2011, 頁 2486–2493.
[6]C. Liu, S. Gong, C. C. Loy及X. Lin, 2012, “Person Re-identification: What Features Are Important?“, 收入 Computer Vision – ECCV 2012. Workshops and Demonstrations, Berlin, Heidelberg, 2012, 頁 391–401.
[7]R. De Maesschalck, D. Jouan-Rimbaud及D. L. Massart, 2000, “The Mahalanobis distance“, Chemom. Intell. Lab. Syst., 卷 50, 期 1, 頁 1–18, 1月 2000.
[8]M. Köstinger, M. Hirzer, P. Wohlhart, P. M. Roth及H. Bischof, 2012, “Large scale metric learning from equivalence constraints“, 收入 2012 IEEE Conference on Computer Vision and Pattern Recognition, 6月 2012, 頁 2288–2295.
[9]P. M. Roth et al., 2014, Person Re-Identification, S. Gong, M. Cristani, S. Yan及C. C. Loy, Springer, London
[10]X. Zhang等, 2018, “AlignedReID: Surpassing Human-Level Performance in Person Re-Identification“. arXiv, 2018年1月31日.
[11]R. R. Varior, B. Shuai, J. Lu, D. Xu及G. Wang, 2016, “A Siamese Long Short-Term Memory Architecture for Human Re-identification“, 收入 Computer Vision – ECCV 2016, Cham, 2016, 頁 135–153.
[12]Q. Xiao, K. Cao, H. Chen, F. Peng及C. Zhang, 2016, “Cross Domain Knowledge Transfer for Person Re-identification“. arXiv, 2016年11月18日.
[13]H. Yao, S. Zhang, R. Hong, Y. Zhang, C. Xu及Q. Tian, 2019, “Deep Representation Learning With Part Loss for Person Re-Identification“, IEEE Trans. Image Process., 卷 28, 期 6, 頁 2860–2871, 6月 2019.
[14]L. Zheng, Y. Huang, H. Lu及Y. Yang, 2019, “Pose-Invariant Embedding for Deep Person Re-Identification“, IEEE Trans. Image Process., 卷 28, 期 9, 頁 4500–4509, 9月 2019.
[15]L. Wei, S. Zhang, H. Yao, W. Gao及Q. Tian, 2017, “GLAD: Global-Local-Alignment Descriptor for Pedestrian Retrieval“, 收入 Proceedings of the 25th ACM international conference on Multimedia, New York, NY, USA, 10月 2017, 頁 420–428.
[16]H. Zhao等, 2017, “Spindle Net: Person Re-Identification With Human Body Region Guided Feature Decomposition and Fusion“, 2017, 頁 1077–1085.
[17]K. He, X. Zhang, S. Ren及J. Sun, 2016, “Deep Residual Learning for Image Recognition“, 2016, 頁 770–778.
[18]A. Hermans, L. Beyer及B. Leibe, 2017, “In Defense of the Triplet Loss for Person Re-Identification“. arXiv, 2017年11月21日.
[19]M. Hirzer, C. Beleznai, P. M. Roth及H. Bischof, 2011, “Person Re-identification by Descriptive and Discriminative Classification“, 收入 Image Analysis, Berlin, Heidelberg, 2011, 頁 91–102.