Study On Robust Moving Object Extraction And Tracking Methods Based On Human Memory Mechanism

Moving object detection and tracking is one of the challenging topics in computervision field at all times due to the complex and changeable real-world scenes, objectappearance changes, and various special circumstances, etc. At present, most of theexisting algorithms can obtain good performances in handling one or several of the issuesin practical applications. However there is no such a moving target extraction and trackingmethod that could handle all kinds of uncertain circumstances and adapt to differentapplications. Therefore, studying on new moving object extraction and tracking algorithmsis still a key issue in the fields of computer vision. The main contributions of thisdissertation are as follows.1) Firstly, the mechanism of human brain memory has been studied thoroughly. Inspiredby the mechanism that human process the memory information, a memory-based cognitivemodel for visual information processing is proposed. According to human three-stagememory model, three memory spaces are defined: ultra-short-term memory space(USTMS), short-term memory space(STMS), and long-term memory space(LTMS), whichare used to store the current and the past information. The proposed model can imitatesome cognitive functions of human brain such as remembering, recall, forgetting, learning,classification and recognition, etc. In addition, some behaviors for manipulating theinformation in memory spaces and the corresponding rules are also defined in the model.2) To handle the scene with sudden changes such as opening or closing a door, amemory-based Gaussian mixture model (MGMM) is proposed based on the proposed memory-based vision information processing model (MVIPM) and Gaussian mixturemodel. In the model, each pixel of every frame is processed and transferred through threememory spaces. During updating, if the background changes, the past backgrounddistributions are marked and stored into the LTMS before they are replaced by a newdistribution. When the similar background occurs, the corresponding distribution can berecalled from the LTMS, which can adapt to the background changes rapidly. Experimentalresults show that the proposed background modeling method outperforms the state-of-artof GMM-based background modeling.3) To handle sudden appearance changes and serious occlusions, a memory-basedtemplate updating (MTU) algorithm is proposed based on the proposed MVIPM. Thetemplate updating model is then incorporated into particle filter (PF) and mean shift (MS)framework for robust object tracking. During tracking, the past templates can be stored inthe memory spaces. Therefore when the similar model appeared again, it can be recalledfrom the memory space to adapt to the variations of the object. Experimental results showthat the proposed tracking framework is robust to sudden appearance changes and heavyocclusions and the performance outperforms that of total template updating (TTU)mechanism based algorithm.4) In order to solve the problem of the loss of diversity among particles, a noveldouble-layer particle filter is proposed. The particles are divided into two layers: the parentparticles and the child particles. The child particles are used to remember the latest state ofthe parent particles and optimize the parent particles. In addition, only the parent particlesare updated during re-sampling while the child particles remain unchanged, whichmaintains the diversity of the particles to some extent. Finally, the parent particles are usedto estimate the state of the object. Experimental results show that the tracking performanceof the proposed double-layer particle filter outperforms that of the basic particle filter.5) Combined MVIPM with multi-agent system (MAS), a memory-based multi-agentsystem for tracking the moving objects is presented. The three-stage human memory mechanism is introduced into a multi-agent co-evolutionary process for finding a bestmatch of the appearance of the object. Each agent can remember, retrieve or forget theappearance of the object through its own memory system by its own experience. A numberof such memory-based agents are randomly distributed nearby the located object regionand then mapped onto a2-D latticelike environment for predicting the new location of theobject by their co-evolutionary behaviors, such as competition, recombination, andmigration. Experimental results show that the proposed method can deal with largeappearance changes and heavy occlusions when tracking a moving object. It can locate thecorrect object after the appearance changed or the occlusion recovered, and outperformsthe traditional particle filter based tracking methods.