Study On Segmentation Algorithms Of Human Object In Video Surveillance

In the field of computer vision, the technique on video segmentation is widelyconsidered as a research hotspot. Scholars at home and abroad have put forward manymethods, but so far, there is no general and reliable method. Especially few methodscan deal with the abrupt illumination and many segmentation results are not integral.Therefore, our study focuses on human object segmentation under the complicatedbackground. Three improved methods about the segmentation of motion human objectand human face are proposed in this paper.Based on the motion characteristic of human object, the paper firstly researcheson Mixture of Gaussians and Bayesian decision in detail and puts forward twoimproved methods. The first one is an improved segmentation algorithm based onMOGs(mixture of Gaussians), where innovation is that it can deal with the abruptillumination. In order to adapt the light abrupt variation, we use a kind of variationdetection method to check the illuminance variation. Then carry on reclassification ofthe results of MOGs method with gradient information, avoiding a great deal ofbackground misclassified as foreground. In second method, a binary map of initialforeground segmentation is achieved by performing Bayesian strategy according tocolor, spatial, and temporal features. In order to obtain more integral and accurateforeground objects, three filters based on mathematical morphology are proposed toprocess the primary segmentation results. According to size filter, some noisebackground patches are removed, while small holes within foreground objects aresewed up by hole filter and the ghosts of human objects are removed through gradientfilter.But these two kinds of method are confined to the motion human objectsegmentation. When the human objects are motionless, these methods will judge them as background wrongly. Owing to this, the paper puts forward a human facesegmentation method based on skin color. The method establishes histogram in HSVcolor space by pixel, and utilizes 3-D affine to model the skin color's transformationfollowing the light variation. A linear compound model combining Markov and Wienerprediction is used to predict skin color's histogram of current frame. Then skin areascan be segmented and the detection of human object can be obtained.Extensive experiments are performed with various video sequences, which provethat these methods are robust and of high segmentation accuracy.