Study Of Cigarette Smoke Recognition In Indoor Video Surveillance

Cigarette smoke is a serious threat on the human body health, the National Health and Family Planning Commission had already submitted the draft to the State Council which is the Regulations that tobacco control in public places.And the challenges faced by cigarette smoke recognition are mainly from the sensitivity of cigarette smoke to environmental conditions, and the limitations of the cigarette smoke video data set in the quantity and quality.Therefore, the research of cigarette smoke recognition in video surveillance, has very important and far-reaching significance.In this paper, we utilize the existing hardware devices to establish cigarette smoke video data sets, analysis of static characteristics and dynamic characteristics of cigarette smoke in detail, and then point out that the statistical features, color layout features and direction features of smoke can effectively describe the characteristics of cigarette smoke; and on this basis, we propose multi-feature fusion method to improve the anti-interference of the feature vector.In order to improve the anti-interference ability and reliability of the feature vectors, this paper proposes the use of feature selection algorithm to enhances the ability of initial feature vector to adapt to the environment.In view of the classic Simba feature selection algorithm that considers all samples have the same importance of the problems and not fully consider the problem of the effect of abnormal values, we propose the introduction of sample weights and diversity measure theory. At the same time, Although the multi feature fusion can effectively improve the ability to adapt to the environment, but the redundant characteristics between features will be affect the final recognition rate. So, this paper puts forward a kind of selection algorithm that based on mutual information(MI_Simba), so as to establish the optimal feature vector of a certain environmental conditions.Through the simulation analysis of the results showed that the MI_Simba feature selection algorithm can effectively improve the environmental adaptability and reliability of the feature vector.Cigarette smoke recognition problem is still at the primary stage, taking intoaccount the less prior knowledge and the special nature of smoke recognition, this paper proposes using the Adaboost algorithm as the basal classifier of cigarette smoke recognition/classification, this is because of Adaboost classifier has the following advantages: strong generalization ability, good stability, excellent detection rate and requiring small priori knowledge. In order to improve the classifier’s recognition efficiency, we introduce the cost sensitive mining theory to enhance the recognition classifier’s pertinence. In addition, taking into account the finiteness of self-built cigarette smoke video data sets, this paper proposed the cost sensitive Adaboost algorithm based on the small sample machine learning, namely the support vector machine.Through the experiment, we verify the effectiveness and superiority of this algorithm in the identification of cigarette smoke.