Visual Object Tracking Based On Correlation Filter

Nowadays,with the development of urbanization,the construction of smart transportation and smart city has a very positive effect on the development of modern city.As a key technology for building smart transportation and smart city,video object tracking plays a crucial role in intelligent surveillance video analysis.Therefore,researching the video object tracking technology can not only help us to analyze surveillance video more quickly and accurately in massive video data,promote the development of intelligent intelligent surveillance video analysis,but also promote the the construction of smart transportation and smart city.The main task of video object tracking technology is to estimate the position and the size scale of the target object in the following frames given a sequence of images or input video and the initial status of the tracking target object.The tracking process generally includes that,first,extracting the feature of the target object according to the previous frame state to construct an expression model of the target object,and applying the model in the current frame.Secondly,employing different methods and strategies to estimate the position and the size scale of the target object,and then updating the model of the target.In recent years,thanks to the development of deep learning and correlation filters,many excellent tracking algorithms have been proposed,which have made the object tracking technology develop rapidly.However,what affects their performance is the various challenges during the tracking process,such as illumination changes,background clutter,bad weathers,and background occlusion,and so on.This thesis focuses on the challenges,and carries out related research on tracking algorithm from the perspective of single modal and multi-modal.The main work and contributions include the following two aspects(1)In order to reduce the impact of feature noise and redundancy,we propose a correlation filter tracking algorithm driven sparse coding.Specially,first,a dictionary is constructed by the clean target feature of the initiate frame,and the feature of the subsequent frame is encoded into more discriminative reconstructed feature by employing sparse coding.However,compared with the original feature,the reconstructed feature may not be able to represent the moving target object completely due to the target appearance changes in the following frames.Therefore,we employ the reconstructed feature and original feature to learn the correlation filters jointly.Thanks to the advantages of joint learning,we propose a unified framework to learn correlation filters and sparse coding at the same time.We conduct experimental comparison analysis on three public data sets to verify the effectiveness and robustness of the proposed algorithm.(2)In order to reduce the impact of illumination changes and bad weathers on object tracking,we propose a cross-modal correlation filters tracking algorithm based on the synergistic low rank constraint.Specially,the infrared camera can capture the infrared radiation emitted by an object whose temperature exceeds absolute zero.Therefore,visible and thermal data can provide strong complementary information,and thus fusing them can boost tracking accuracy.Afterwards,it is found that different modalities have similar correlation filter to make them have consistent localization of the target object through the experiments.So considering the interdependence between visible and thermal modalities,we exploit the low-rank constraint to learn correlation filters collaboratively to promote the fusion of two modalities effectively,thereby enhancing the tracking performance.