Research On Recognition Algorithm For Smoking Behavior Based On Multi-feature Of Smoke

Smoking is a behavior common in our daily life, whose harm has gradually beenknown by the public. Most public places at present are non-smoking, especially shoppingmalls, railway stations in which a relatively large number of people frequently flow.Compared with the great efforts to tobacco control which are increasing, monitoringmethods for smoking behavior are still relatively under-developed. In this paper, themethod with regard to smoking behavior recognition based on multi-feature is studied.This method takes advantage of the analysis to visual information, with smoke detectioncombined with the smoking behavior characteristics detection to determine whether thesmoking action occurs. The manpower required in tobacco control will be effectivelyreduced and the efficiency will be improved through this method.This paper studies smoking behavior recognition from two aspects separately, the firstis to determine the region of smoke and the second is to extract feature vectors, both ofwhich are the keys to this issue. In this paper, the spatio-temporal interest regions aredetermined within the video sequences through smoking behavior firstly. Utilize HSVcolor model effectively to describe the translucency peculiar to tobacco smoke, and thentake advantage of different detection methods and thresholds aimed at backgrounds withdifferent color characteristics to mark out suspected smoke pixels in the spatio-temporalinterest regions. Furthermore, the smoke regions suspected are obtained throughmorphological operations. In order to exclude the non-smoke regions existed in thesuspected smoke regions, the suspected smoke regions are analyzed and judged by thesmoke dynamic characteristics. The characteristics of area change and centroid trajectoryof the suspected smoke region are analyzed in the process of diffusion and flowing ofsmoke, and finally feature vectors can be obtained through the description to thesecharacteristics using the feature descriptor.Establish a video sample library through recording video, and figure out the numberof smoke samples and non-smoke samples. Then the feature vectors are extracted andnormalized. Finally, one classifier is trained by training set whose performance is tested by testing set then. After many experiments the kernel function of the classifier is chosen andthe validity of the feature vectors is verified. The final results of recognition accuracyreach more than90percent.