本論文基於Haar-like特徵的Adaboost演算法對車用影像進行行人、二輪車及汽車之多重物件偵測，訓練樣本的過程中，結合行人、二輪車及汽車三者樣本組成一個多重物件偵測分類器，樣本優化上採OpenCV提供的Performance程序，保留分類器命中的樣本並去除遺漏的樣本，重新訓練新的分類器，提升分類器的偵測準確率且降低誤判率，此作法能提供性能較佳的分類器並改善25%的分類器性能調校時間。對於分類器的偵測結果、鏡頭架設及環境因素三方的配合，本論文採用基準線(Focus line)及投票演算法(Voting algorithm)對偵測結果施予調校與穩定，其中，Focus line的功能為確立物件出沒的區域並將偵測影像中的過上及過下的極端型誤判加以消除，而基於偵測較穩定的分類器結合Focus line的偵測率表現可使誤判率從26.51%下降到3.52%，有效地修正整體的誤判程度；在偵測影像中，觀察出標示圖框閃爍且不穩定的情況，此現象為偵測率及誤判率影響因素之一，對此本論文提出投票演算法，儲存前後一至二張的偵測資訊後，透過鄰近矩陣將偵測結果加以分群並在群集中定義繪製標示圖框的門檻。在實作過程中，針對Current frame置後(3張)及Current frame置後(5張)結合不同的圖框繪製門檻，進行實驗與評估結果，由於Current frame置後的因素，使無法取得整體所需的資訊，令偵測率及誤判率皆明顯攀升且誤判率抑制不易的問題，為此本論文將Current frame置中，可同時取得前後張一至兩張的偵測資訊並判斷Current frame之間的前後連貫的關係，來決定是否繪製Current frame的位置。在偵測性能評估數據中，鏡頭設置於車輛的水箱罩上所偵測時，偵測率從91.47%至91.76%，誤判率從26.51%下修至3.33%；而鏡頭為置設置於車廂內時，偵測率從96.51%至96.61%，誤判率從24.86%下修至22.62%；由於鏡頭架設高度的因素，鏡頭架設較低的感興趣區域(ROI)較鏡頭架設高度大，以致Focus line較能發揮效用，而投票演算法在不同架設環境中皆能發揮其功能。

This thesis describes the results of experiments in which Adaboost algorithm was mixed with Focus line and Voting algorithm. The proposed technique could be useful in multiple object detection. It is possible that combining pedestrian, motorcycle and vehicle made one classifier to detect. The Optimization of sample depends on OpenCV Performance program. We save the sample which was hit. This way improve the accuracy of detection and reduce the false alarm rate and save 25 % of the classifier performance tuning time. For the detection result classifier lens erection and with environmental factors tripartite this thesis Focus line and voting algorithm Voting algorithm to detect and adjust the results lend stability, which, Focus line of function for establishing infested area and detect objects in the image over and over on the extreme type misjudged be eliminated under, and based classifier to detect more stable binding Focus line performance make false detection rate increased from 26.51% dropped to 3.52 percent, effective correction overall level misjudgment; in detecting images, observe the situation marked frame flashes and unstable, this phenomenon detection rate and false positive rate of one of the factors, which make this paper voting algorithm to detect after one or two before storing information sheets through neighboring matrix will detect clustering results to be plotted and labeled in the cluster definition frame threshold. In the process implemented for the Current frame set after (3) and Current frame set (5) a combination of different frame drawing threshold, conduct experiments and evaluate the results, due to factors Current frame set after that can not be achieved overall the necessary information to make the detection rate and false positive rate and false positive rate were to rise significantly inhibit difficult problem, for this thesis Current frame set in, which can get around one to two information sheets to detect and determine Current frame coherent relationship between Current frame to determine whether the position drawn. In the detection performance evaluation data, the lens is set at the time of the vehicle detected by the tank cover, the detection rate from 91.47 to 91.76 percent, from 26.51 percent error rate was revised down to 3.33%; while the lens is set to set in train compartments , the detection rate from 96.51 to 96.61 percent, from 24.86 percent false positive rate was revised down to 22.62%; erection due to the height of the lens elements, the lens erected lower region of interest (ROI) than the height of the lens erection that Focus line more able to be effective, while voting algorithms in different environments also erected to function.

誌謝 i
摘要 ii
Abstract iii
目錄 iv
圖目錄 vi
表目錄 viii
第一章 緒論 1
1.1 研究動機 1
1.2 研究背景 1
1.3 研究貢獻 2
1.3.1 多重物件偵測 2
1.3.1 基準線(Focus line) 2
1.3.1 投票演算法(Voting) 2
1.4 論文章節架構 2
第二章 文獻探討 3
2.1 Adaboost演算法 3
2.1.1 Haar-like特徵計算 4
2.1.1.1 Haar-like特徵的(u, v)條件 4
2.1.1.2 特徵左上座標A(x1,y1)、右下座標B(x2,y2)的範圍 4
2.1.1.3 計算特徵數量 4
2.1.1.4 特徵數量總和 5
2.1.2積分圖像計算 6
2.2 基準線 8
2.2.1 基於車輛高度基準線 (horizon lines) 8
2.2.2 基於鏡頭架設與消失點基準線 8
2.3 鄰接矩陣 9
第三章 演算法探討與開發 11
3.1 Adaboost多重物件偵測之分類器 11
3.1.1 整體樣本及局部樣本的選擇 11
3.1.1.1 行人樣本擷取 12
3.1.1.2 二輪車樣本擷取 12
3.1.1.3 車輛樣本擷取 14
3.2 環境相異之樣本調和 16
3.3 基準線(Focus line) 19
3.4 投票演算法(Voting algorithm) 21
3.4.1 宣告一個一維陣列P 22
3.4.2 匯集多重物件偵測結果 22
3.4.3 將陣列P導入鄰近矩陣W 23
3.4.4 鄰近矩陣之資料掃瞄與擷取 24
3.4.5 鄰近矩陣之計算量 25
3.4.6 鄰近矩陣之座標點分群 26
3.4.7 鄰近矩陣W回傳群集序號至陣列P 28
3.4.8 標示各個群集的位置座標 28
第四章 實作與實驗結果 30
4.1軟硬體設備需求及參數設定 30
4.2分類器內容 31
4.3基準線(Focus line) 33
4.4投票演算法(Voting algorithm) 35
第五章 效能評估 37
5.1 性能評估方式 37
5.1.1 偵測率評估 37
5.1.2 遺漏率評估 38
5.1.3 誤判率評估 39
5.2 分類器性能及群集分布分析 39
5.2.1 向前參考兩張Frame之效能評估 40
5.2.2 向前參考四張Frame之效能評估 41
5.2.3 數值統整 42
第六章 結論 46
第七章 未來展望 47
參考文獻 48

[1].P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” in Proceedings IEEE Computer Vision and Pattern Recognition, CVPR, pp. 511-518, 2001.
[2].G.R. Rakate, S.R Borhade, P.S Jadhav, M.S Shah, “Advanced Pedestrian Detection system using combination of Haar-like features, Adaboost algorithm and Edgelet-Shapelet”. IEEE International Computational Intelligence &; Computing Research Conf. ICCIC, Coimbatore, pp. 1-5, Dec. 2012.
[3].W. Chunpeng, D. Lijuan, M. Jun, F. Faming, W. Xuebin, “Detection of Front-View Vehicle with Occlusions Using AdaBoost,” in Proc. Int. Conf. Information Engineering Computer Science, Wuhan, pp. 1-4, 19-20 Dec. 2009.
[4].D. Bobo, L. Wei, F. Pengyu, Y. Chunyang, W. Xuezhi, Y. Huai, “Real-time on-road vehicle and motorcycle detection using a single camera” IEEE International Industrial Technology. Gippsland, VIC, pp. 1-6, Feb. 2009.
[5].X. Guoqing, W. Xiaocui, L. Li, W. Zhengbin, “Real-time pedestrian detection based on edge factor and Histogram of Oriented Gradient”. IEEE International Information and Automation Conf. ICIA, Shenzhen, pp. 384-389 , June 2011.
[6].H. Roncancio, A.C. Hernandes, M. Becker, “Vision-based system for pedestrian recognition using a tuned SVM classifier”. IEEE International Engineering Applications Conf. WEA, Bogota, pp. 1-6, May 2012.
[7].Z. Ning, Y. Qixiang, J. Jianbin, “Fast pedestrain detection with cascade classifiers”. IEEE International Information, Computing and Telecommunication Conf, Beijing, pp. 550 -553, Sept. 2009.
[8].Z. Fan, L. Guizhong, W. Xing, “An efficient AdaBoost tracking algorithm based on the particle framework”. IEEE International Signal Processing, Communications and Computing Conf. ICSPCC, Xi&;#39;an, pp. 1-4, Sept. 2011.
[9].C. Wen-Chang, J. Ding-Mao, “A cascade classifier using Adaboost algorithm and support vector machine for pedestrian detection”. IEEE International Systems, Man, and Cybernetics. SMC, Anchorage, AK, pp.1430-1435, Oct. 2011.
[10].K. Lertniphonphan, S. Aramvith, T.H. Chalidabhongse, “Human action recognition using direction histograms of optical flow”. IEEE International Communications and Information Technologies Conf. ISCIT, Hangzhou, pp. 574-579, Oct. 2011.
[11].L. Bo, Y. Qingming, W. Kunfeng, “A review on vision-based pedestrian detection in intelligent transportation systems”. IEEE International Networking, Sensing and Control Conf. ICNSC, Beijing, pp.393-398, April 2012.
[12].T. Kowsari, S.S. Beauchemin, J.Cho. “Real-time Vehicle Detection and Tracking Using Stereo Vision and Multi-View AdaBoost”. IEEE International Intelligent Transportation Systems. ITSC, Washington, DC, pp.1255-1260, Oct. 2011.
[13].郭志宏(2010)。單眼視覺的行人偵測與追蹤(碩士論文)。國立中央大學資訊工程研究所，桃園。
[14].F. Zhao, L.C. Jiao, H. Liu, X.B. Gao, “A novel fuzzy clustering algorithm with nonlocal adaptive spatial constraint for image segmentation”. Signal Processing, 91(4), pp.988-999, 2011.
[15].S.C. Chen, D.Q. Zhang, “Robust image segmentation using FCM with spatial constraints based on new kernel-induced distance measure.” IEEE International Systems Man and Cybernetics Part B: Cybernetics, 34(4): 1907-1916, Aug. 2004.
[16].R.C. Luo, L. Hsin-Hung, “Design and implementation of efficient hardware solution based sub-window architecture of Haar classifiers for real-time detection of face biometrics”. IEEE International Mechatronics and Automation Conf. ICMA, Xi&;#39;an, pp.1563-1568, Aug. 2010.
[17].S. Esmaeeli, I. Gholampour, “Reduced memory requirement in hardware implementation of SVM classifiers”. IEEE International Electrical Engineering Conf. ICEE, Tehran, pp. 4648-4653, June 2006.
[18].M. Gressmann, G. Palm, O. Lohlein, “Surround view pedestrian detection using heterogeneous classifier cascades”. IEEE International Intelligent Transportation Systems Conf. ITSC, Washington, pp. 1317-1324, Oct. 2011.
[19].A. Prioletti, A. Mogelmose, P.Grisleri, M.M. Trivedi, A. Broggi, T.B. Moeslund, “Part-Based Pedestrian Detection and Feature-Based Tracking for Driver Assistance: Real-Time, Robust Algorithms, and Evaluation”. IEEE International Intelligent Transportation Systems, vol. 14, no. 3, pp 1346-1359, Sept. 2013.