Research On Correlation Filter Visual Tracking Based On Feature Representation And Model Optimization

Visual object tracking has been one of the key topics of research in the field of computer vision.It has been widely used in a variety of real-world scenarios,such as intelligent transportation systems,intelligent surveillance systems,human-computer interaction,unmanned aerial vehicles,and robotics.An object tracking algorithm aims to model the appearance of the target using the given initial information,overcome the interference of complex factors such as target occlusion,fast motion,and illumination changes in subsequent image sequences,and finally achieve effective inference of the target position and size.For the only information available about the initial appearance of the target,how to fully and effectively mine it is very helpful to improve the performance of a tracker.From well-designed lightweight handcrafted features to complex but more discriminative deep features,researchers have presented a large number of excellent methods and theories for target representation.However,there are still some problems to be solved.Firstly,traditional feature fusion methods usually use fixed weights to fuse different feature layers to achieve information interaction and sharing,but they cannot automatically adjust the fusion weights according to the changes in the scene,which lacks sufficient flexibility and robustness.Secondly,the fused features are often characterized by high dimensionality and channel diversity,which seriously affects the accuracy and real-time of tracking.It is still a very challenging problem to identify discriminative channels from a large number of feature channels and eliminate redundant channels.Moreover,how to utilize regularization to constrain and guide the filter model while ensuring the real-time performance of the tracker remains to be further addressed.In view of the above issues,this paper explores the feature fusion,feature selection,regularization,and model learning based on the correlation filter tracking framework.The main works of this paper are as follows:(1)A multi-complementary feature adaptive fusion visual tracking algorithm based on game theory is proposed.Firstly,two robust combined features are obtained by combining features with complementary properties selected from handcrafted features and deep features,respectively.Secondly,the idea of game theory is adopted to treat the two combined features as two sides of the game.Through continuous games in the tracking process,the two achieve the best dynamic balance,so as to obtain a more robust fusion feature.Finally,the proposed tracker is extensively evaluated with other related algorithms on the OTB2015 dataset.The experimental results show that our tracker is more suitable for achieving robust visual tracking in complex scenes such as occlusion and deformation.(2)A robust visual tracking algorithm via adaptive feature channel selection is proposed.Our method adaptively selects discriminative feature channels for target representation by calculating the energy relationship between the background and the foreground in each feature channel.In addition,an adaptive model updating strategy is used to alleviate the problem of model degradation caused by incorrect model updating.Finally,a large number of experiments are carried out on five popular tracking benchmark datasets,including OTB2015,TC128,UAV123,VOT2018,and La SOT.The experimental results demonstrate the effectiveness and robustness of the proposed algorithm.(3)A real-time UAV visual tracking algorithm with joint learning of background suppression and target perception is proposed.First,a mask is generated using a priori knowledge to construct a filter with target perception and a filter with background suppression,respectively.The background suppression filter is implemented by introducing a context-aware regularization term into the traditional correlation filter formula,which can suppress background noise and further mitigate unwanted boundary effects.The target perception filter only models the target patch,and the robustness of the model to target occlusion and deformation is enhanced by introducing a temporal regularization term.The two are jointly optimized by the ADMM algorithm to mutually constrain each other in the training stage and complement each other in the detection stage.Experimental results on the UAVDT and DTB70 dataset verify the merits of the proposed method.Furthermore,the speed of our tracker is about 26 fps,which can meet the real-time requirements of UAV tracking.