Research On Object Tracking Based On Adaptive Spatial Regularized Correlation Filters

In recent years,visual object tracking as a key technology in computer vision and image processing,has a wide range of applications in video surveillance,human-computer interaction,smart transportation,and so on.The subject attracting a large number of outstanding experts and scholars to devote in research and achieve considerable results.However,the unknown external environment and variable personal factors bring many unpredictable challenges to the research of visual object tracking,such as deformation and similar object occlusion.In the face of the expanding market demand,the precision and speed of the tracker has become an evaluation metric that cannot be ignored in algorithm design.Therefore,it has become crucial to design a method that can cope with the multiple challenges mentioned above and still achieve real-time tracking.This thesis aims to construct a robust object tracking models to ensure that the models can well incorporate the proposed mechanisms and ultimately achieve precision real-time tracking of the object.Two main aspects of research are carried out in this thesis.Firstly,a background-aware correlation filter algorithm based on adaptive saliency regularization was proposed.Due to the dynamic change of the search region,the periodicity hypothesis,although solving the problem of insufficient training samples,inevitably introduces boundary effects,which can lead to severe failures in the detection stage.First,our tracker adds a background penalty factor into the correlation filter,and proposes a novel spatial regularization by using the saliency detection.Secondly,in order to solve the model better and faster,we introduce a novel energy function to solve the spatial weight,and apply the alternating direction method of multipliers method to deduce the closed-form solutions of each subproblem of the objective function efficiently.Thirdly,we propose an adaptive updating mechanism based on the variation of target appearance and the reliability of tracking results,which can update the model online by adjusting the spatial weight distribution for precisely tracking in the spatiotemporal domain.Fourthly,two models are constructed to estimate the position and the scale of the target respectively.One model adopts hand-crafted features at multiple scales to select the optimal scale,while the other model predicts the optimal position by fusing hand-crafted features with deep features extracted from the trained network models.Finally,extensive experiments are carried out on the tracking datasets to validate the performance of our tracker.Secondly,a noise-aware correlation filter algorithm based on dynamic regression label was proposed.Typically,discriminative correlation filter-based trackers utilize a learning regressor to achieve object and background distinctions,where the regressor regressing cyclic samples into a fixed target label.In addition,such trackers introduce fixed regularizations into the model to improve robustness in learning of target appearance.However,both a fixed target regression label and a fixed regularization may cause tracker incompatibility for unknown aerial scenes,resulting in inaccurate tracking.In this algorithm,we make full use of the background information reflected in the generated response maps during the detection phase.Firstly,the local maximum of the response map of the previous frame is applied to distinguish noise,and introduced into the regression learning phase to suppress noise effectively,thus achieving dynamic changes in the object regression labels.Secondly,the local response map variation of the successive frames is introduced into the spatial weight to construct a new spatial regularization;meanwhile,the global response map variation is introduced into the constraint coefficients of the spatio regularization to achieve update together with correlation filters.Finally,for the newly established object model,we adopt the alternating directional multiplier method to better and faster solve the closed solutions of each sub-problem of the object loss function,and conduct extensive experiments on three unmanned aerial vehicle datasets.