Study On Shadow Detection Of Moving Object For Video

With the rapid development of social economy and city construction, the sharpexpansion of urban population is putting great pressure to public security. In recent years,along with the advance of digital image processing techniques, intelligent video surveillancesystem becomes more and more popular, which plays a crucial role in the construction ofintelligent city, safe city and smart city. In the field of computer vision, the core of intelligentvideo surveillance system is object detection, which is the basis of video content analysisincluding scene analysis, processing and behavior understanding. However, the study ofobject detection still confronts many challenges due to the complexity of video scenes.Shadow is one of the optical phenomena in nature as well as a kind of common imagedegradation phenomenon, which possesses two visual features similar to object. One is thatshadow and background have obvious differences; the other is that shadow and object havethe same motion characteristic. These two properties make it easy to be detected as object.Shadow reduces the accuracy of object detection, which may cause object merging, objectshape distortion and even object loss. It can severely impact the subsequent video contentanalysis. Therefore, shadow detection has become the key problem for video surveillancesystem, which has important theoretical significance and extensive application value.On the basis of optical shadow model, the study on moving shadow detection is carriedout deeply and four different shadow detection methods are put forward. The effectivenessand superiority of the proposed methods are demonstrated by extensive experiments onseveral standard benchmarks and comparisons with several state-of-the-art methods. Themain work is summarized as follows:1. When the color information of object is close to background, the object is usuallydetected as shadow mistakenly. To overcome the drawback of single color feature, we presenta moving shadow detection method by combining color with texture in terms of the textureconsistency. Firstly, we adopt the intensity and chromaticity in HSV color space to detectshadows in foreground image. Meanwhile, we utilize Local Binary Patterns (LBP) and localvariance to describe the texture for foreground image, calculate texture similarity betweenforeground and background for detection. Subsequently, logical operation is used to combinethe two results. The proposed method takes the color constancy and texture consistency ofshadow into account simultaneously, which makes the two properties complement with eachother. Experimental results indicate that our method achieves better detection accuracy.2. Most of conventional shadow detection methods based on multiple features use eachfeature independently in serial mode and then adopt yes/no logical mechanism to determine whether a foreground pixel is shadow or not. Different from these methods, we address amultiple features fusion method by the utilization of intensity, color and texture in parallelmode. Besides the intensity, color features are firstly extracted from multiple color spaces andmulti-scale images while texture features are obtained by entropy and LBP from multiplechannels. Then, normalize each feature map and fuse these maps to generate the final map bydecision level fusion strategies. Finally, we employ appropriate threshold to final feature mapfor classification. Experiments on various benchmarks validate that the proposed method withlinear weighted fusion is superior to other fusion strategies and some representative methods.3. To address the sensitivity against noise or uncertain factors for pixel-based shadowdetection methods, we suggest an adaptive region segmentation method. Two differentmethods are exploited to segment foreground. One is Affinity Propagation (AP), and the otheris watershed algorithm. The former is to partition the foreground into blocks withoutoverlapping, extract color feature from blocks and then use AP to cluster for pre-segmentation;the latter perform watershed algorithm to gradient image of the foreground and several subareas are obtained according to gradient changes. Then, we compare gradient changes in eachsub area, compute texture similarity between sub areas in consecutive frames and in currentframe and background for classification. Obviously, the method utilizes the texture similarityin both of intra-frame and interframe simultaneously, meanwhile carries out the combinationof spatial and temporal features. Compared with the method of fixed block, experimentalresults demonstrate that the adaptive region segmentation method can retain the consistencyproperty in sub area well. Yet, it also has better robustness than pixels-based methods.4. The effectiveness of most existing shadow detection methods relies on somerestrictive assumptions and illumination changes in scenes, and even requires each video tohave a fixed parameter set. However, it is not adaptive for complex scenes. Consequently, wepropose a moving shadow detection method based on statistical discriminant model. First,different types of features are extracted from man-made labeled pixels and formed a featurevector as original samples. The number of features is the sample dimension. Second, partialleast square (PLS) is applied to dimensional reduction and logistic discrimination (LD) isadopted for classification. Meanwhile, the statistical discrimination model PLS-LD isestablished. Finally, PLS-LD is used to classify new input pixel. The method does not dependon any threshold and can automatically judge the class of pixels. Extensive experimentalresults and cross training-test on various videos justify the effectiveness and generalizationability of proposed method.