摘要
近年來國人常見的慢性疾病如：癌症、腦血管疾病、心臟疾病、呼吸系統疾 病等都與吸菸有重大關係。自從政府於民國98 年實施新菸害防制法，明訂包括 室內公共場所、室內三人以上工作場所及大眾運輸工具內全面禁止吸菸。「菸」 不只對吸菸者本身健康造成危害，燃燒不完全的二手菸甚至比一手菸對人體健康 傷害更大。基於維護及保障多數非吸菸者的健康權。因此，本研究提出一套運用 多特徵融合進行吸菸行為偵測系統。 所提吸菸行為偵測系統主要包括人臉、膚色、香菸及煙霧偵測等四要項。首 先在人臉偵測方面，利用Haar-Like 特徵及AdaBoost 演算法來判斷影像中是否有 人臉存在。而膚色偵測方面，使用正規化RGB 並計算出可能是膚色的像素。在 香菸偵測方面，找出影像中飽和度較低且亮度較高的部分，且其大小小於膚色部 分，並與膚色重疊，則視為香菸的位置，接著計算出香菸與人臉的比值，作為吸 菸行為偵測的第一個特徵。在煙霧偵測方面，則是針對不同時間下的臉部範圍做 煙霧顏色判定，並計算其變化量，作為吸菸行為偵測的第二個特徵。最後利用最 近鄰居分類法並根據前述兩特徵，判斷影像中是否有吸菸的行為。同時若在連續 十張影像中出現七次以上吸菸行為，則判定該段影像視訊有吸菸行為發生。
並列摘要
In recent years, it is common that citizens in Taiwan have chronic diseases such as cancer, cerebrovascular, cardiac, respiratory disease, etc, which are critically related with smoking. Since government implemented the new Tobacco Hazards Prevention Act in 2009, it indicates that smoking is not permitted in the place, such as indoor public places, the indoor workplace with three or more workers, and in public transportation. "Smoking" is not only hazardous to the health of the smokers himself, but also impose greater health problem to the secondhand smokers because of incomplete combustion of the tobacco. In order to protect the rights of the majority people, who do not smoke, while reducing extra burden from manual monitoring, this study proposes a simple but effective smoking behavior detection system by using feature fusion.. The smoking behavior detection system includes four components: face, skin, cigarette and smoke detections. First, the Haar-like features and AdaBoost algorithm are used to detect the faces in an image. The skin colors are detected by using normalized color pixels. The cigarette in the image is detected with low saturation and high brightness, overlapping area with skin area, and smaller size than that of skin. The ratio of cigarette to face is then calculated as first feature for smoking event detection. In addition, the amount of smoke color change in face area at different time is calculated as the second feature. The k-Nearest Neighbor rule (KNN) in terms of the two mentioned features is used to determine whether smoking behavior exists in the image. Finally, if there are more than six times of smoking events detected among ten consecutive images, the video clip is claimed smoke behavior event occurring.