Robust Visual Tracking Via Low-Rank Representation

Visual object tracking is a significant computer vision task which can be applied to many domains such as visual surveillance,motion analysis and human-computer interaction.It has been attracting attention over the decades due to its great value in theory as well as in application.Despite extensive research on this topic,it still suffers from difficulties in handling complex object appearance changes cause by factors such as illumination changes,partial occlusions and pose changes.In order to enhance the robustness,I proposed a new algorithm after analyzing the benefits and drawbacks of the visual object tracking methods based on low-rank matrix representation.The contribution of my work are the following:1.I presented a new appearance model.To fix the defect of the existing algorithm based on low-rank representation towards partial occlusion,I proposed a repeated lowrank representation method,which exploits both global appearance information and local appearance information by dividing the entire appearance model into two parts called global low-rank representation and local low-rank representation respectively.Also,the max pooling feature is used in my method to alignment the coefficients of both global representation and local representation,while background templates are used to avoid drift.2.I redefined the rules of computing the particles’ weight.Because the rule of computing the particles’ weight was too simple to release the impact of partial occlusion,a new rule was made to compute not only global weight but also local weight.With the help of local weight,more partial information is taken into consideration while computing the final weight of particles,which can enhance the robustness towards occlusion.While computing the local weight,a judgement of occlusion is made to further reduce the impact of occlusion.3.I generated a new strategy of dictionary updating.A cumulative probability sequence based on quadratic function is used to represent the probability of the updating template set.This sequence not only ensures that the information of template is real-time,but also maintains the diversity of templates.To avoid the wrong information of the object brought in the dictionary while occlusion occurs,a judgement is made.And the different strategies would be taken for the different results of the judgement on whether occlusion occurs.