行人穿越道行人監控系統之研究

Abstract

使用影像序列作即時的行人偵測被廣泛地應用在智慧型交通系統(intelligent transportation system, ITS)，例如交通監控以及交通號誌控制，目前存在許多行人偵測的方法，但是隨著行人數目以及行人交錯程度的增加，行人偵測的效能將會變的很差。根據行人穿越道上行人的特性，我們提出一個方法可以有效地偵測及追蹤行人穿越道上在人群交錯狀況下的行人。 一般而言，行人會在行人穿越道的兩端穿越道路，因此我們只在行人穿越道的兩端偵測行人，一旦行人被偵測到便開始追蹤行人，如此我們可以得知行人在任意時間的位置並且減少運算複雜度；另外，我們將攝影機鏡頭設置成與道路方向垂直，這樣的好處是行人的影像將會左右對稱，這個特性會被用來辨識行人。行人偵測的方法是先偵測出類似人頭外形的邊緣(edge)，接著使用結合許多行人特性的模糊推論系統(fuzzy inference system, FIS)判斷偵測到的邊緣是否可能是人頭；在某些情形下人頭的邊緣並不明顯，以上的方法將會失效，但對於孤立的行人可以利用行人的輪廓作為偵測行人的依據。追蹤的方法是比對顏色的直方圖(histogram)，預測行進路徑的方法是使用簡易的無限脈衝響應(infinite impulse response, IIR) 濾波器。實驗結果顯示，當光線不是太暗而且行人穿越道不是太長下，本論文所提出的方法對於行人穿越道的行人偵測及追蹤有不錯的效果。

Real-time pedestrian detection using video sequences has myriad applications in the intelligent transportation system such as traffic monitoring and traffic signal control. Many methods have been proposed for the detection of pedestrian. However, as the number of pedestrian and the amount of their occlusion increases, the system performance degrades. According to the properties of pedestrians at crosswalks we propose an approach to efficiently detect and track pedestrians in the crowded situation at crosswalks. Generally, the pedestrian will pass through the road from one side to another side of the crosswalk. Therefore we only detect pedestrians on the both sides. Once a pedestrian has been detected, the system will start to track the pedestrian. Then we can know the location of the pedestrian at any time, and the computation complexity can be reduced. In addition, we set up the camera with the direction perpendicular to the road. In this way an advantage is that the pedestrian's image will be about symmetrical. This characteristic will be used for recognizing the pedestrian. For detecting pedestrians the system searches human head shape in the edge image first and then uses the fuzzy inference system which combines several pedestrian’s characteristics to judge whether it is human head. That in some situations edges are not apparent makes the above method inactive but the isolated pedestrian can be detected by human shape contour. For tracking pedestrians the system exploit the color histogram match. The next position of every pedestrian can be predicted by infinite impulse response filter. The experimental results show that if the light is not too dim and the crosswalk is not to long, then the proposed method has good performance on detecting and tracking pedestrians.