Adapting Object Appearance Changes For Robust Visual Tracking

Object tracking is an important topic in computer vision research,and has applied in military and civil fields in recent years.Despite great progress has been made,a reliable object tracking method still needs to cope with many challenges such as deformation,scale variation and occlusion.This thesis designed a robust discriminative online tracking method based on correlation filter,the highlight of which is the ability to adapt appearance changes.Towards the complicated and changeable real-world scene,a Multi-channel Selected Correlation Filter(MSCF)is proposed,through which the holistic appearance based on PCA-HOG is modeled by learning short-term dynamic features.The long-term stable appearance is represented by ORB keypoints,while maintaining a feature set that is employed to retrieve the missing tracking.In order to combine the similarity evaluation scores of aforementioned two appearance features,a confidence fusion framework is incorporated to obtain the final output confidence map.Particularly,the position of partially occluded object is located by a local keypoints voting mechanism.Furthermore,the scale and rotation of target,which can contribute to the further analysis of higher level semantic information,is acquired via a status estimation method based on structural consistency constraint.To alleviate error accumulation caused by inappropriate learning and promote the stability and sustainability of the tracking process,an adaptive learning ratio is set to update the filter templates for each channel of the MSCF independently.The results of plenty of qualitative and quantitative experiments on the OTB-100 benchmark suggests that the proposed tracker outperforms several state-of-the-art methods.