Object Tracking Based On Sparse Appearance Representation

As a hot topic and technology in the computer vision field,visual object tracking plays an important role in video surveillance systems,human-computer interaction,video searching,video compression and military,etc.However,in practice it is hard to track the visual object effectively when suffering various interferences,including occlusion,illumination change,background clutter and scale,and pose of visual object.Based on the above issues,we integrate the sparse representation theory into the particle filter framework,and use the positive image patches voting and multi-task learning theory to improve the stability of the tracking method in dealing with the influence of interference factors.The detail contents of this thesis are as follows:(1)We introduce the background and significance of visual object tracking based on sparse object tracking,and summarize the existing problems in visual object tracking task.Through the tracking experiment on the challenging data sets of various standards,we analyze the advantages and limitations of the tracking method.(2)We integrate the Multi-Task learning into the visual object tracking based on the assumption that few but similar templates are required to represent all particles.We consider the representation of each particle is a single task,and employ the lp,q mixed norms to enforce joint sparsity and share the learning process of all the particles.Finally,we use an Accelerated Proximal Gradient(APG)method to achieve the fast convergence of the sparse solution.Experimental results demonstrate that the proposed approach achieves the stable tracking result in background clutters,occlusion and illumination change sequence.(3)L1 tracking method only considers the global information of the template,which cannot well track the object with the similar background,we propose a tracking method based on positive patch voting to overcome the problem of drifting.By establishing the confidence function and the similarity function of the frame,we calculate the weights of positive patches and vote the optimal location.The experimental results demonstrate that the proposed approach can deal with background clutters,occlusion,illumination change,and fast movement.