Research On Real-time Visual Tracking Algorithm In Complex Scenes

Visual tracking is one of the most challenging computer vision problems,and it is widely developed in numerous applications,such as video surveillance,drone tracking,self-driving vehicle,human computer interaction,auxiliary medical diagnosis and national defence security.In generic,tracking task is to estimate the trajectory of an arbitrary target in an image sequence,given only its initial location.Owing to the high tracking speed and good performance,correlation filter(CF)based trackers have drawn considerable attention.The high performance attributes to the useful characters of the circulant matrix,which makes it possible to transfer the process of correlation to the fourier domain and augment the samples implicitly.Therefore,based on CF and the CF-based tracker,we improve and develop our trackers.One successful tracker should satisfy robustness and real-time,simultaneously.However,the variation and the complexity of tracking environment always results in unexpected dilemma.Through intensive analysis of various tracking algorithms in recent years,we conclude that the dilemma is caused by two aspects: the onefold traditional features and the failure to re-detect after the target loss results in low robustness;the framework of tracker is too complex to satisfy real-time tracking speed.To address the problems mentioned above,we improve the tracking algorithms from two aspects: the feature model and the overall framework of the tracker,meeting both the high robustness and tracking speed.In terms of the improvement of the feature model,we proposed an adaptive object tracking approach with complementary models based on template and statistical appearance models,named CCMT.Both of these models are unified via our novel combination strategy.The proposed approach embraces the advantage of both models,and it can efficiently handle the various and complex scenarios such as occlusion,illumination variation or background clutter.In addition,we introduce an efficient update scheme.With the priori threshold,the update scheme can efficiently avoid model contamination,and thus improve the performance of our approach further.The experimental results demonstrate that our proposed approach CCMT is far more robustness than KCF in terms of robustness,while achieving superior performance at speed that far exceed the real-time requirement.To develop a novel framework,we proposed a multi-level CF-based tracking approach named MLCFT.Different from traditional methods,MLCFT applies convolutional neural network(CNN)for extracting deep features,which provide strong capacity for target representation.To take more spatial information from shallow layers and semantic information from deep layers into consideration,we utilize multiple layers of CNN to represent the target,simultaneously.And an effective fusion method based on relative entropy is introduced to combine the complementarty features from different layers of CNN.Meanwhile,we further explore the potential capacity of CF with two-stage detection: primal detection and cascaded re-detection.Through a pipeline operation,including threshold suppression,regional restriction and non-maximum suppression mechanism,the process of redetection becomes more efficient and accurate.Moreover,a novel model update strategy is proposed,which enhances the tracking performance further.Several ablation experimental results verify that our design choice in MLCFT is correct and feasible,MLCFT can deal with more various scenarios such as fast motion,motion blur,occlusion,background clutter or low resolution,etc.Additionally,the quantitative and qualitative experimental results also demonstrate that our proposed approach MLCFT outperforms the most state-of-the-art trackers,while operating at a real-time speed.