Structural Correlation Filter Visual Tracking Based On Feature Fusion

Visual Object Tracking(VOT)is one of the most core and challenging tasks in computer vision research and application which is widely used in the fields of manmachine interface,monitoring and security system,traffic monitoring and control,medical diagnosis and so on.It is also an important component of some advanced research topics in computer vision,such as scene detection and behavior recognition.The apparent characteristics of the object such as color,shape,texture are one of the important research topics in the visual object tracking.Many scholars begin with the object modeling when they study the method of visual object tracking.In this thesis,in order to establish a robust object tracking method,the study of VOT method is not limited to the way of object modeling,the VOT system is divided into motion model,feature extraction,observer model,model update and post-processing based on the commonly used visual object tracking methods.The each functional part are studied and improved from the view point of visual object tracking.The main work of the thesis have the following points:(1)On the VOT Benchmark data set of CVPR2016,the common methods of five components in VOT system and their influences on video track performance are studied by contrast experiment.On this basis,the design idea of structured correlation filtering(FFBS-CF)based on multi-feature fusion is proposed.Firstly,the target will be split and the loose space relationship of every parts will be applied to the tracker.Then the feature fusion framework is proposed to integrate the overall color statistical features of the tracking target into the tracker,using the component space relation and the target overall color statistical characteristics of the combination of target tracking.(2)Based on the above improvements,the FFBS-CF method is designed and implemented.The feasibility and validity of the method are verified on the CVT2016 VOT Benchmark test data set.It is compared with the MOSSE,KCF and DSST algorithms in the specific case of the tracking results.FFBS-CF tracker performance is better than MOSSE method in the whole;When occlusion,motion blur,apparent deformation and rapid motion occur,the performance is slightly better than that of KCF and DSST methods;When the light changes,the L component of the color template in the FFBS-CF tracker is distorted,so its tracking performance is not as good as the KCF and DSST methods.The FFBS-CF target search box is fixed,the performance of the FFBS-CF target is significantly reduced when the target scale is drastically changed by introducing excess background or missing part of the target.(3)The FFBS-CF method is applied to the trace module of the ITS system.The tracking performance is compared with the MOSSE,KCF and DSST algorithms.The results show that the FFBS-CF method enhances the overall adaptability of the system to complex environments.However,when the apparent feature of target is split by occlusions,the performance is reduced.Therefore,the FFBS-CF method still has a large optimization and improvement space.