在本論文中,我們提出一個以俯視視訊為基礎之行人計數系統。在本系統中,使用俯視視訊來代替傳統的紅外線或柵門進行行人計數,最大的優點是不影響行人通行,且因為行人區塊不易重疊,而使計數效果更為準確。系統架構主要由行人偵測以及行人計數兩方面所組成:在行人偵測方面,應用以K-mean為基礎的影像分割技術來擷取行人區塊,可以使系統應用較不受光線所影響;在行人計數方面,採用graph matching技術,同時整合區塊本身及區塊間的特徵進行追蹤,使追蹤結果較可信賴,最後根據區塊的移動路徑以及區塊面積修正計數的結果。最後的實驗結果顯示,本系統在光線不斷變化或行人通過數量較多時,仍可以進行有效的計數。
In this thesis, a people counting system based on top-view video sequences is proposed. This system consists of foreground people detection and people counting algorithm. For people detection, an image segmentation method based on k-means clustering is employed to extract human figures. In order to use in different of illumination conditions, we use region merging to remove shadows of each object. With this approach, our system can be applied in outdoor environments. In the people counting part, human regions are tracked and counted based on a graph matching algorithm. Tracking results are used to determine the direction of region movement based on unary and binary features. Our system has been tested in many different cases of pedestrian density. We give examples of the system counting people in real-time in describe.