Research On Correlation Filter Based Visual Tracking Algorithms

Visual tracking is a fundamental topic in computer vision and has numerous applications in video surveillance,human-computer interaction.Given the bounding box in the first frame,its task is to estimate the location of a target in an image sequence.It is challenging in complex scenes,e.g.,deformation,occlusion,illumination variation.Significant progress has been made in recent years.Correlation filter based visual tracking algorithms have been very popular due to the excellent performance.However,the robustness is influenced by the scene.Therefore,this thesis focuses on correlation filter based visual tracking algorithms to improve the performance on deformation and illumination.The main contributions are as follows:(1)Feature integration with adaptive importance maps for correlation filter based visual tracking is proposed.HOG features focus on gradients of the target,which can describe the target's shape and are robust to illumination,while color features can extract features from homogeneous regions and are robust to deformation.To take into consideration importance and complementary information of features,the newly designed algorithm jointly learns filters,as well as importance maps in each frame.These importance maps utilize the advantages of different features,in order to improve the robustness to deformation and illumination.Meanwhile,for each feature,an importance map is shared by its all channels to avoid overfitting.In addition,the newly designed algorithm introduces a regularization term for the importance maps and uses the penalty factor to control the significance of features.Experimental results validate the effectiveness of importance maps and show that the newly designed algorithm achieves competitive accuracy.(2)Low rank correlation filter for visual tracking is introduced.Correlation filter based tracking algorithms depend on multi-channel features to achieve superior performance.To reduce the redundancy of multi-channel features,the newly designed algorithm introduces the constraint about low rank filters.Therefore,multi-channel features can be represented in the subspace,in order to improve the robustness to the change in target's appearance.For effectively optimizing the new algorithm,this thesis employs alternating direction method of multipliers to solve low rank filters.Experiments demonstrate that the newly designed algorithm provides competitive accuracy compared to several trackers.(3)Distractor-aware correlation filter for visual tracking is presented.Cyclic Shifts are the key of correlation filter.However,they lead to the inaccurate representation and boundary effects,which degrade the performance of the tracking algorithm.To reduce boundary effects,we utilize color histograms to find the distractor,e.g.,the similar object,clutter background.For making full use of hard negative samples to strengthen the discriminative ability of the model,this thesis learns filters which have a low response to the distractor,in order to improve the robustness to deformation and avoid overfitting.Experimental results demonstrate that the newly designed algorithm achieves competitive accuracy.