Research On Vision Object Tracking Based On Local Sparse Coding

Visual object tracking is an important application in computer vision.Combined with the rapid development of computer vision technologies,such as image processing,face detection,and so on,the technology of visual object tracking has made great progress in the past decades.However,visual object tracking is still challenging for the complex application environments,such as illumination variation,partial occlusion,background clutter,and so on.Generally,a visual object tracking model includes several parts,such as appearance model,motion model,model updating,and so on.Based on the technology of local sparse coding,this dissertation proposes several improvements and innovations on visual object tracking technology..The main works and innovations of this dissertation include three parts.(1)Base on the theory of the local sparse coding,this dissertation proposes a dual-scale structural local sparse appearance model.In the existing researches on visual object tracking based on sparse coding,sparse representation is used either to capture the global features of the target or to capture the local features of the target.Thus,this dissertation proposes a dual-scale local structure sparse appearance model to capture both the quasi global features and the local structure features of the target.The dual-scale appearance model is consist of two different scale local sparse coding models,where the local sparse coding appearance model with large scale is used to capture the quasi global structure information of the target and the local sparse coding appearance model with small scale is used to capture the local structure information of the target.The dual-scale local sparse coding appearance model takes into account both the quasi global structure information and the local structure information of the target,therefore it can locates the target more accurately and deal with the problem partial occlusion of the target better.In addition,this dissertation proposes a new joint mechanism to combine these two structural local sparse coding appearance models.(2)Based on the technologies of kernel correlation filter and sparse coding,this dissertation proposes a hybrid tracking framework based on the kernel correlation filter and the local sparse coding.The technology of visual object tracking based on kernel correlation filtering has achieved great success in recent years.However,there are still some problems in visual target tracking based on correlation filter,such as appearance model of the target,scale variation of the target,and so on.This dissertation proposes a hybrid tracking framework based on the kernel correlation filter and the local sparse coding to deal with the problem of target scale variation.First,the tracking model based on the kernel correlation filter is used to locate the target preliminarily;Then,the tracking model based on local sparse coding and particle filtering is used to locate the target further and capture the scale variation of the target;Finally,the accurately position of the target and the scale information of the target are used to be trained for the kernel correlation filter model.The kernel correlation filter and the local sparse coding model in our hybrid tracking framework can complementary each other.The kernel correlation filter model has good accuracy and efficiency,but it lacks the processing of target scale variation.The sparse coding model can capture the scale variation and deal with the partial occlusion of the target better,but it has poor accuracy relative to the tracking technology based on discriminative model.In our hybrid tracking framwork,the kernel correlation filter model can gives the local sparse coding model an accurate object initial position,and the local sparse coding model provides an accurate object position and the scale variation information of the target for the kernel correlation filter model.(3)Based on the popular motion model of particle filtering,this dissertation proposes a remarkable local resampling technique based on particle filtering.Generally,an ideal outcome for particle filtering model needs a large number of particles to approximate the posterior.Therefore,scholars extract remarkable particles from all particle set by non-linear function in recent years.The remarkable particles can reduce the number of particles and can improve the accuracy of particle filter model.However,when these non-linear functions are used to extract remarkable particles,the optimal local particles will be weakened or even neglected.Thus,this dissertation proposes a new resampling method to extract locally remarkable particles.Two parts of remarkable local particles are extracted by a weight threshold and a distance threshold to track and analysis the target.According to the remarkable local resampling technique and a global transport model,this dissertation proposes a locally remarkable particle filtering model.