Correlation Filter Architecture Based Robust Visual Tracking Algorithm

Visual tracking is to estimate the complete trajectory of the visual object in the image sequence.Only the initial frame object position is known during the tracking process without other prior knowledge.Usually the tracking system consists of initial position,appearance model,motion estimation model,and object positioning model.Because visual tracking has a large number of applications in the fields of robotics,autonomous driving,intelligent monitoring,augmented reality and camera autofocus,it has become the most important research topic in the field of artificial intelligence and computer vision,and involves multi-disciplinary cross-fusion,which has attracted the attention of many scholars.Although there are some excellent visual tracking algorithms,it is still challenging to perform accurate and robust visual tracking in complex scenes.Mainly due to the existence of illumination variation,scale variation,occlusion,deformation,motion blur,fast motion,in-plane rotation,out-of-plane rotation,out-of-view,background clutter,and low resolution in scenes.As they will cause serious appearance change,the performance improvement is bottlenecked.Based on this background,this thesis studies the robust visual tracking algorithm based on correlation filters.The research contents and contributions of this thesis are as follows:First,a visual tracking algorithm based on correlation filter and candidate samples screening strategy is proposed.The algorithm defines a large search area by radius to fully capture the appearance information left by the occluded target,which effectively avoids difficult to handle the appearance change caused by the occlusion in the case of sliding window local search.The candidate sample screening strategy selects candidate samples that may be target region from a large number of samples cropped from the larger search area,and then correlation filter is used to calculate the spatial correlation response of these samples to estimate the target position.The experimental results show that the proposed algorithm can effectively deal with the drift caused by occlusion scenes,and can also deal with challenging scenes such as fast motion,background clutter,in-plane rotation,deformation,motion blur,out-of-plane rotation,and out-of-view.Second,a visual tracking algorithm coupled with local and global correlation filter models is proposed.When occlusion occurs,the motion estimation model constructed by a single correlation filter tends to be sensitive,showing tracker drift and not being able to recover normally,and then appearing the phenomenon of tracking failure.According to the internal complete structure of the target itself,this paper adds sparse constraint between the local and global correlation filter models to form a coupled model.The model can tolerate the outliers of the local model and use the appearance information of the remaining visible regions when the target is occluded,thus alleviating the drift risk of the tracker and improving the robustness of tracking.In addition to being able to deal with occlusion problem,the algorithm can also cope with challenging scenes such as illumination variation,scale variation,motion blur,out-of-plane rotation,and background clutter.Third,an adaptive visual tracking algorithm based on correlation filter and hierarchical deep feature is proposed.The bottom layer of the deep features has the high resolution characteristics of the hand-crafted features,as well as the high-level semantic features,showing strong invariance for serious appearance change,and the ability to improve the robustness of visual tracking.In this thesis,the five layers of feature information of the bottom layer,middle layer and high layer are used to express the appearance of the target respectively.Each layer learns a correlation filter.By calculating the correlation response of different layers,the mapping position of different layers can be obtained,and then the target location can be obtained by adaptively weighting the mapped locations of the different layers.Experimental comparison with the state-of-the-art tracking algorithms show that the algorithm significantly improves the accuracy and robustness by using the five layers of deep features of the bottom layer,middle layer and high layer.Finally,a visual tracking algorithm based on interactive fusion hierarchical deep feature strategy is proposed.The hierarchical deep feature mutual fusion strategy is used to make each layer feature mutually reinforcing and characteristic sharing,so as to achieve a correlation filter visual tracking framework with both accuracy and robustness.Using bilinear interpolation,the five-layer visual feature maps of the bottom,middle,and high-level are adjusted to a fixed size by up-sampling,and then the five-layer deep feature maps are interactively merged through two stages to construct an appearance model that shares texture and semantic information.The target position can be effectively estimated by the motion estimation model.The experimental results show that the proposed algorithm uses the five-layer deep features to construct a visual appearance model through the two-phase interaction fusion,which achieves an effective unity of accuracy and robustness,and the overlap success rate is improved to some extent.