Shape Information Based Moving Object Detection

In the intelligent video surveillance,moving object detection as the basis of gait recognition and behavior analysis,plays a key role.Background subtraction,as one of the moving object detection methods,is greatly influenced by the environment.The detected results are often not complete objects.In order to make the detection results close to real targets,the top-down shape information can be fused into the background subtraction methods.If the background subtraction method does not consider shape information,a target may be divided into multiple connected domains in dynamic(such as leaf dithering,fountains,etc.)and camouflage scenes,which affects high-level applications such as gait recognition.In order to solve this problem,how to improve the result of background subtraction by shape information is studied.The result is rebuilded to a relatively complete object by on-line shape prior image and off-line training shape model.The main contributions of this dissertation are summarized as follows:(1)The background subtraction algorithms are studied.A algorithm of embedding the shape prior information into the Markov random fields(MRF)post-processing is presented.The local energy function corresponding to the shape prior information is designed.The past results of current frame are used to guide the follwing moving object detection.In addition,moving object is located by connected domain analysis to increase the probablity that the moving object is labelled as foreground.The MRF post-processing with shape prior imformation is used to modify the results based on the Gaussian Mixture Model(GMM)background subtraction.Experimental results show that the proposed method outperforms GMM and GMM with MRF post-processing results,which can improve the extraction of moving objects in complex scenes,especially in scenes with dynamic interference.(2)In this dissertation,a method that DBM is applied to background subtraction post-processing is proposed.This method can slove the problom that one object is divided into many connected domain in camouflage scenes.DBM is applied to establish the model of pedestrian shape.Experiments show that this method can rebuild the shape of human body,and can obtain complete shape in real camouflage scenes.(3)The DBM background subtraction method is applied to gait recognition and classification.The gait recognition system is sensitive to foreground image contours.When the shape image is occluded,the recognition rate of gait is decreased.The experimental results show that the occluded object is rebuilded by DBM,and the gait recognition rate is improved.This shows that there is a possibility that DBM may learn each person's personalized shape features.