臺灣博碩士論文加值系統

現今各國社會經濟快速的發展，私人交通工具數量也日益增多，除了造成交通壅塞的情形，更使得交通意外事故頻傳，而台灣地狹人稠，機車是最簡便輕巧的交通工具，因此在一般道路上最為常見，但相對的機車的肇事率最為高，也容易與人發生擦撞，因此本論文提出一套能快速有效偵測車周的機車騎士及行人的偵測及追蹤演算法以降低意外的發生。
在物件偵測的領域中最常被選作為描述物件特徵的描述器為Histogram of Oriented Gradients(HOG)，因為HOG是在圖像的局部區塊進行特徵抽取，所以對圖像幾何及光學形變都能保持很好的不變性，但在複雜的環境下，還是可能有物件誤判的情形。
傳統HOG的正規化的範圍為單一的Block，但有可能忽略周圍區塊所帶來的影響，導致影像中出現較複雜的紋理時容易有誤判的情形，且因為傳統HOG的特徵維度極大，使得運算速度難以達到即時的效果，因此，本論文提出一套能加速及改良正規化範圍的HOG演算法。此外，亦提出複合式特徵檢測，利用HOG判斷機車騎士的安全帽是否在偵測物件中出現，以降低誤判率。
本論文提出一套新的有效偵測區域的定義方式，能有效濾除影像中不必要偵測的部分以加速偵測時間。為了使本論文應用於更多情境，將演算法實作於嵌入式平台NVIDIA TX2中，整體偵測率於嵌入式系統及個人電腦均達到9成以上。

In recent years, the fast development of society and economy in various countries, the number of private transportation vehicle is also increasing, not only causes traffic congestion, but also cause traffic accidents more and more frequently. Because Taiwan is a small island, which is highly populated, so the motorcycle are the most convenient transportation, also the most common on the road. Oppositely, the motorcyclists’ traffic accidents rate is highest in all transportation and also be prone to accidents with pedestrian. Therefore, this thesis proposes a surrounding pedestrian and motorcyclist detecting and tracking system with vehicle algorithms to decrease accident rate.
The most commonly used descriptor for describing object features in the field of object detection is Histogram of Oriented Gradients (HOG). Because HOG performs feature extraction in local blocks of the image, the geometric and optical deformation of the image can be kept very invariant, but in a complicated environment, there may be cases of misjudgment of objects.
The traditional HOGs normalization scope is a single block, but it is possible to ignore the influence of the surrounding blocks, it’s liable to have a misjudgment when encountering complicational textures in image, and because the feature dimension of traditional HOG is very large, the computational speed is difficult to achieve in real time. Therefore, this thesis proposes a HOG algorithms that can accelerate and improve the normalization range. In addition, a compound feature detection method is also proposed to determine whether the helmet is appearing in detected motorcyclist object to reduce the false positive rate. This thesis propose a new ROI define method, which can effectively filter out unnecessary parts of the image to speed up detection time. In order to apply this thesis to more situations, the algorithm is implemented in the embedded platform NVIDIA TX2, and the overall detection rate is over 90% in both embedded systems and personal computers.

摘 要 i
ABSTRACT ii
誌謝 iv
目 錄 v
表目錄 vii
圖目錄 viii
第一章　緒論 1
1.1 研究背景與動機 1
1.2 研究目的 2
1.3 論文架構 3
第二章 文獻回顧 4
2.1 方向梯度直方圖 4
2.1.1 Cascade Histogram of Gradients 10
2.1.2 強化式HOG特徵 13
2.2 支援向量機Support Vector Machine 16
2.3 感興趣區域 18
2.3.1 改良式透視法快速Window篩選 18
2.3.2 車輪偵測 20
2.4 安全帽偵測 23
2.5 物件追蹤 25
2.5.1 Template Matching 25
2.5.2 Kernelized Correlation Filter 26
第三章 研究方法 28
3.1 系統架構 28
3.2 模型訓練 30
3.2.1 訓練樣本蒐集 30
3.2.2 影像灰階化及Gamma校正 32
3.2.3 加速型強化式HOG特徵抽取 33
3.2.4 SVM樣本訓練 38
3.3 物件偵測 39
3.3.1 改良式可變視窗有效偵測範圍機制 41
3.3.2 複合式特徵檢測 44
第四章 實驗結果與分析 47
4.1 實驗環境 47
4.2 實驗結果 49
4.2.1 個人電腦平台實驗結果 50
4.2.2 嵌入式平台實驗結果 53
4.3 文獻比較 56
4.3.1 強化式HOG計算效能比較 56
4.3.2 OpenCV HOG效能比較 57
4.3.3 有效偵測範圍比較 59
4.3.4 安全帽偵測演算法比較 62
4.4 實驗總結 64
第五章 結論與未來工作 65
5.1 結論 65
5.2 未來工作 66
參考文獻 67

[1] 內政部警政署-交通部統計查詢網-中華民國交通部A1類道路交通事故
[2] Viola, Paul, and Michael Jones, "Rapid object detection using a boosted cascade of simple features," Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on. Vol. 1. IEEE, 2001.
[3] N. Dalal and B. Triggs, "Histograms of oriented gradients for human detection," IEEE Computer Society Conference, Computer Vision and Pattern Recognition, vol. 1, pp. 886-893, 2005.
[4] Wikipedia, “Support Vector Machine” https://en.wikipedia.org/wiki/Support_vector_machine
[5] Learn OpenCV, ”Histogram of Oriented Gradients” https://www.learnopencv.com/histogram-of-oriented-gradients/
[6] Wikipedia, ” Histogram of oriented gradients” https://en.wikipedia.org/wiki/Histogram_of_oriented_gradients
[7] Zhu, Qiang, et al, "Fast human detection using a cascade of histograms of oriented gradients," Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on. Vol. 2. IEEE, 2006.
[8] T.K. An and M.H. Kim, “A new Diverse AdaBoost classifier,” IEEE International Conference, Artificial Intelligence and Computational Intelligence, vol.1, pp.359-363, 2010.
[9] Wang, Xianji, et al, "Palmprint identification using boosting local binary pattern," Pattern Recognition, 2006. ICPR 2006. 18th International Conference On. Vol. 3. IEEE, 2006.
[10] 成明，基於強化式HOG特徵與快速物件分類技術之嵌入式快速行人物件偵測及追蹤系統，碩士論文，國立臺北科技大學資訊工程系研究所，臺北，2015。
[11] Ma, Guanglin, et al, "Vision-based pedestrian detection-reliable pedestrian candidate detection by combining ipm and a 1d profile," Intelligent Transportation Systems Conference, 2007. ITSC 2007. IEEE. IEEE, 2007.
[12] 陳昱彰，基於HOG特徵之車周物件偵測及追蹤系統，碩士論文，國立臺北科技大學資訊工程系研究所，臺北，2016。
[13] Pathasu Doungmala and Katanyoo Klubsuwan, ” Half and Full Helmet Wearing Detection in Thailand using Haar Like Feature and Circle Hough Transform on Image Processing,” IEEE International Conference on Computer and Information Technology,2016.
[14] Henriques, João F., et al, "High-speed tracking with kernelized correlation filters," IEEE Transactions on Pattern Analysis and Machine Intelligence 37.3 (2015): 583-596.
[15] INRIA, “INRIA Person Dataset” http://pascal.inrialpes.fr/data/human/
[16] 維基百科, "交叉驗證", https://en.wikipedia.org/wiki/Cross-validation_(statistics)
[17] Wikipedia,“Pyramid (image processing) ” , https://en.wikipedia.org/wiki/Pyramid_(image_processing)
[18] EDOM TECHNOLOGY ,“TX2 Module”, https://www.edom.com.tw/tw/product
[19] Logitech, “C922 PRO STREAM CAMERA”, https://www.logitech.com/zh-tw/product/c922-pro-stream-webcam
[20] OpenCV3.2, “OpenCV”, https://github.com/opencv/opencv