A Hierarchical Shadow Detection Method Used For Traffic Video

In recent years, along with the advance of digital image processing techniques,intelligent video surveillance system becomes more and more popular, which plays a crucialrole in the construction of intelligent city, safe city and smart city. In the field of computervision, object detection is the basis of video content analysis including scene analysis,processing and behavior understanding. However, the research of object detection stillconfronts many challenges due to the complexity of video surveillance scenes.Shadow is one of the optical phenomena in nature as well as a kind of common dropphenomenon, which possesses two visual features similar to objects. One is that shadow andbackground have obvious differences, the other is that shadow and object have the samemotion characteristic. These two properties make it easy to be detected shadow as object. Theexisting of shadow reduces the accuracy of object detection, and may cause object merging,object shape distortion and even object loss. It will give rise to serious influence forsubsequent video content analysis. Therefore, shadow detection has become an inevitable keyproblem for intelligent video surveillance system, which has important theoreticalsignificance and extensive application value.This paper presents a hierarchical shadow detection method used in intelligent trafficsurveillance videos. First, the foreground is obtained by using Gaussian Mixture Model(GMM), and shadow detection is based on the result of foreground detection. Then, in therough detection stage, we construct shadow dictionary and object dictionary by using trainingsamples, the reconstruction error of foreground regions is adopted to determine whether theregion is shadow or not. Meanwhile, brightness feature is utilized to separate shadows fromforeground regions. The fusion of the two results improves the accuracy of shadow detection.In the refined detection stage, features derived from illumination invariant are applied torefine the rough detection result and the final detection result is obtained after spaceadjustment.The proposed method is simulation experimented on four databases, and compared withsome well-known methods. The four database are named as Video I, Video II, Video III andVideo IV，according to the distance between camera and the road. Experimental results showthat proposed method could detect the moving shadow accurately in traffic surveillance videoscenes, and the proposed method is better than the contrasts.