Research On Long-term Real-time Object Tracking Algorithm Based On Kernelized Correlation Filters

Computer vision based object tracking refers to the process of continuously locating objects in continuous video sequences,it has urgent application requirements in many fields such as human-computer interaction,security monitoring,augmented reality,medical imaging and autonomous navigation.In recent years,with the increasing complexity of practical application scenarios,the challenges of object tracking algorithms are also increasing.Illumination variation,scale and shape variation,fast motion,motion blurring,occlusion,out-of-view and real-time maintenance all pose great challenges to the robust tracking of objects in videos.Therefore,designing and implementing real-time,robust object tracking algorithms for complex environments has high research and application value.In this thesis,the problem of single-object long-time tracking of unmanned mobile platform is achieved by integrating FHOG and Color Names features,introducing occlusion judgment,learning rate dynamic adaptation and feature dimension reduction method,and using separate scale filters.The learning rate adaptive kernelized correlation filter algorithm have stronger ability of object appearance representation,better ability to cope with object scale variation,more reasonable template update strategy,and higher real-time performance.On this basis,a multi-filter kernelized correlation filter algorithm with detector is designed and implemented by combining short-term memory filter,scale filter,long-term memory filter and online SVMbased re-detector,so that it not only can quickly determine the object position and scale,but also can retain the historical appearance "memory" of the object,re-search the object when the filter template is inaccurate or the object is occluded to improve the robustness of the object tracking.In order to further solve the problem of tracking failure caused by occlusion on mobile platform,this thesis also designs an SVR-based trajectory prediction module,which can accurately predict the object trajectory after severe occlusion,which significantly reduces the probability of tracking failure of the mobile platform in encountering such a situation.Focusing on the difficulty of long-time tracking of single object on unmanned mobile platform,the kernelized correlation filters algorithm is deeply studied in this thesis.In addtion,the two proposed object tracking algorithms are tested and analyzed on OTB data sets from qualitative and quantitative perspectives respectively,and the results are compared with some traditional excellent algorithms.The experimental results show that the two algorithms proposed in this thesis both have strong competitiveness.They also show that the combination of multi-feature fusion,scale and learning rate adaptive,long-term tracking and some other methods can significantly improve robustness of kernelized correlation filters algorithm under various adverse situations such as the partial and complete occlusion,shape variation,scale variation and illumination variation.These improvements provide a more solid algorithm foundation for the application research of long-term single object tracking on unmanned mobile platforms.