People Counting System In Complex Scenario

Getting the number information of people is of great significance in the area ofcommercial information collection and security monitoring. The traditional manualway of surveillance is very time consuming and tedious. Firstly, it has much moreburden on human resources. What’s more important is that it’s inevitable for a man’smistake after long time’s tedious work. So the automatic, computer vision-basedpeople counting system is drawing more and more attention.Before the computer vision-based people counting method emerged, thereexisted many researches on the density of people. A lot of methods on densityestimation have been proposed. As techniques’ progressing, methods focusing on theexact number of people come out. But there are still a lot of difficulties when they areapplied to real scenarios. Among them are complexity of background, occlusions ofpeople, changes of illumination, and the interference of non-human moving objectscombined with human bodies, etc.This paper focuses on the people counting problem in complex outdoorenvironments. There’re a lot of difficulties in the test video, such as complexbackground, occlusions, and the unconventional body combined with objects such asbicycles and umbrellas. All these factors impose great challenges on people countingsystems. To deal with the problems, a set of local features are extracted according tothe real scenario. And some measures are taken to solve interference. At last, thenumber of people is obtained via the SVRusing the feature vector as its input.Partial occlusion, the interference of some common non-people objects, andsegmentation fault caused by complex background are solved in this way. Meanwhile,since the resolution of modern surveillance video is becoming higher and higher, animproved mixture of gaussian model is used to improve efficiency and make itapplicable for real-time systems.