Study On The Moving Object Detection And Tracking Based On Human Memory Mechanism

Nearly all the existing moving object detection and tracking algorithms have limitations in practical applications.Due to the complex and changeable real-world scenes,the existing methods can achieve good performances when handling one or several of the issues under particular circumstance.Compared to machine vision,the human being have a more powerful visual perception system which can easily recognize and track the interesting target even under the complex scenes.Memory mechanism plays an important role in the procedure of visual perception.With the help of memory system,we can identify the interesting target using parts of feature information and adjust our cognitive behavior according to the logistic knowledge stored in memory.In this paper,we introduce the three-stage memory mechanism model into the existing object detection and tracking approaches in order to improve the robustness under the complex and changeable scenes.The main work of this thesis is listed as follows:1.Inspired by human memory mechanism,a memory-based Codebook model(MCB)is proposed.MCB has the ability to learn the background online that can avoid the problems of false detection and incomplete detection caused by offline training of the traditional background modeling.2.A multi-feature fusion model under the three-stage memory mechanism is proposed for target modeling.The FAST corner feature is fused with the color feature while the temporal information of these features is introduced into the proposed model whose reliability and adaptability to the variation of scenes is improved.3.To solve the degeneration problem of particle filter tracking,a mixed propagation model for particles is proposed.The target's center oriented transition and corner oriented transition are added into the original state transition equation so that the particle's degeneration can be avoided with no loss of diversity.In turn,the extraction area of corners is constraint by the particles' distribution to decrease the quantity of false matching corners.Experimental results demonstrate the validity of the proposed algorithms mentioned above.