隨著科技日益增進，和社會及個人安全的考量，基於影像資訊發展的監控系統成為現今非常重要的研究標的；其中，基於視覺影像的群眾切割，更是廣泛地應用在不同應用系統當中，不論是多人追踨、無人駕駛智慧車、行人偵測、人數統計或是行為分析，都需要準確地的標記出群眾中各個個體的位置，才能進一步作其偵測與分析。 群眾切割是一個極富挑戰性的題目，由於人的肢體動作變化性大，人體的正面和側面輪廓有所不同，每個人的衣著打扮各有不同，再加上一群人走在一起常有彼此遮蔽的問題，更是造成特徵擷取的困難。故我們在本研究中提出一套適用在複雜情境的群眾切割系統，且賦予切割的目標一個信任等級，可被其他應用系統當作參考，例如多人追蹤，有助於提升其偵測和分析結果。 本研究的第一部分，特徵擷取方面利用有向性梯度直方圖HOG (Histogram of Oriented Gradients, HOG) 配合Curvelet (曲波)，配合分類器向量支撐機 (Support Vector Machine, SVM) 對群眾中的個體進行偵測，和單純使用HOG特徵的結果比較，可顯著地提升群眾切割的準確率。第二部分，為了避免人與人之間的遮蔽問題，我們提出一個新的特徵---改進式類Haar-like特徵運算式的有向性梯度直方圖 (Modified Haar of Oriented Gradients, MHOOG)，分別為MHOOG-U和MHOOG-L，以較少的維度資訊，對人體上半身和下半身做準確的偵測，並賦予其信任等級，我們可藉由多張frame之間的信任等級，提供進階研究作較精確的參考。實驗部分呈現本研究所提出的方法可以達成可信度高的群眾切割，在賦予信任等級的偵測實驗中，等級1的每張frame平均的false alarm偵測人數約為0.0227~0.0401，等級4的每張frame平均miss的偵測人數約為0.125~0.4417。本研究確實提供擁有高可靠性參考資訊的群眾切割。 Crowd Segmentation plays an important role in security surveillance systems. We propose a confidence level computation mechanism and three effective feature extraction methods which facilitate crowd segmentation. With confidence level information of multiple frames, segment individuals in crowds can be performed more accurately. Besides, our system can be exploited in subsequent applications such as tracking or human activity analysis. Firstly, we design a new feature descriptor based on HOG (Histogram of Oriented Gradients) and curvelet. Then, we exploit the descriptor and support vector machine (SVM) to detect the full body of the pedestrian. Secondly, we construct an upper body detector and a lower body detector using new features MHOOG-U (Modified Haar of Oriented Gradients for Upper Body) and MHOOG-L (Modified Haar of Oriented Gradients for Lower Body). To locate the upper body and lower body of the individuals in the crowd, MHOOG-U and MHOOG-L are not only reliable and require much fewer dimension information compared with HOG features. Afterwards, we evaluate the segmentation results according to the detection responses of these three parts. According to the experimental results, with quite few the false alarm number per frame about 0.0227~0.0401 in confidence level 1 and low miss number per frame around 0.125~0.4417 in level 4, we confirmedly provide the valuable information from our confidence level computation mechanism. With multiple frames‟ information provided from our system, reliable detection information is evaluated and crowd segmentation can be achieved more accurately.
