Research On Efficient Method Of Detecting Moving Foreground And Abnormality In Road Surveillance Video

Video surveillance is widely demanded in modern society, and there are great amount of applications in both public and private fields. Especially in public places, video surveillance needs massive devices deployments and human resources. Comparatively, an intelligent surveillance system might have following advantages: 24 hour continuous surveillance, high efficiency and robustness, low costs, and protecting private information. In this thesis, two fundamental problems of automatic surveillance are studied, i.e. foreground detection and abnormality detection. There is an assumption that surveillance cameras are static to capture the scene of streets or roads with vehicles and pedestrians.Among foreground detection methods, background subtraction is really popular in real-time surveillance systems. In this thesis, sigma-delta filter is introduced to achieve background modeling. Then, improvements are proposed to better maintain background and balance the update rate and dirty rate. To overcome the problem of slow initialization, median filter is used to establish the initial background. Some morphological operations are taken to fill holes, eliminate tiny noise regions, connect and separate regions, and finally extract smoothing object regions.Among abnormal event detection, features based methods skip the steps of object recognition and tracking, which leads to lower complexity and system demands and has adaptability to distributed in-road surveillance systems. A set of object features are used in this thesis, including shape features, area feature and velocity features. Using these feature vector inputs, classifiers are established to determine the labels of object class, object location, object velocity and finally the abnormality of objects. Based on statistical analysis, we independently optimized these classifiers, and obtain the map of active region and the map of velocity distribution.The proposed foreground detection and abnormality detection schemes are tested on public datasets. Experimental results demonstrate that these schemes can be used in multi-cam real-time surveillance systems.