Research On Robust Visual Object Tracking Algorithm

Visual object tracking technology is one of the core issues of machine learning and also a hotpot of current research.The application of visual object tracking is helpful for accurately obtaining the nature of dynamic appearance of visual object,and for meeting the military and civilian requirements,such as visual navigation,intelligent monitoring,unmanned autonomous system,robot operation and interactive servo,etc.However,the factors like deformation,rotation,occlusion and background clutter seriously influence the accuracy stability during tracking.Since 2010,the correlation filtering tracking which is based on correlation matching in classical signal processing has drawn much attentions of many researchers.Compared with the previous tracking algorithms,correlation filtering tracking has a super high speed and favorable robustness,therefore has been a research hotspot in visual tracking community in recent years,and many excellent correlation filtering tracking algorithms have been proposed successively and get favorable tracking results.Although the performance of current correlation filtering tracking has been improved significantly,due to the complexity of tracking scenario and the existing of multiple challenges,designing a robust and accurate tracking algorithm remains an opening problem.With this researching background,considering the decreasing of the accuracy of the target appearance representation and the weakness against background distractors of current correlation filtering tracking in the complex scenarios,this dissertation combines the correlation filtering tracking with local-global layered coupling visual model and spatialtemporal context information perception,and with the local-global tracking collaboration,multi-feature response fusion,spatial-temporal information joint learning and spatialtemporal feature learning,this dissertation researches how to improve the accuracy in target appearance representation and the ability against the complex background distractors of correlation filters,and also proposes the related tracking algorithms.The main researching works and innovations of this dissertation are summarized as follows.(1)Aiming at the issue that the correlation filter tracking algorithm has insufficient perception for the target appearance itself,which results in its poor adaptability to sudden target occlusion or deformation,a local-global layered coupling visual model is proposed.Combining the layered vision model with the correlation filter tracking framework significantly improves the perceptibility of correlation filter tracking to the target appearance,and the adaptability of correlation filter tracking to complex tracking scenarios.Based on the PSR value of the correlation response and the motion trajectory analysis of the local-global layers,the algorithm studied the trackability,motion consistency of the local patches,and the accurate estimation of the target position of each patch.Then,the accurate estimation of the target position is achieved by selecting reliable patches.The experimental results on video sequences with scale change,rotation,occlusion and deformation as the main challenges show that the proposed tracking algorithm improves the basic KCF algorithm by 10.1% in distance precision and by 7.8% in success rate.(2)Aiming at the issue that the accuracy of the filter model is degraded of existing correlation filter tracking algorithm when handling complex scenes,based on the complementarity of multiple features,a multi-feature response fusion tracking strategy via local-global layered coupling visual model is proposed to improve the adaptability of correlation filtering tracking to complex scenes.The local layer and the global layer adopt different tracking algorithms.The global layer adopts the correlation filtering tracking based on HOG feature,while the local layer studies the trackability of each patch based on the foregroundbackground feature discrimination method,and implements a structured local color model block tracking algorithm SLC.Then,an online SVM detector is introduced to realize the conditional fusion of global and local tracking results in the response maps level,and the final tracking algorithm LGCm F is constructed.The experimental results on OTB2015 dataset show the effectiveness and the superiority of the proposed block tracker SLC and the final tracker LGCm F.Especially,LGCm F improves the state-of-the-art tracker Staple by5.7% and 3.1% in distance precision and success rate respectively,verifying the effectiveness of the proposed tracking strategies further.(3)Considering the issue of single positive training sample,low quality of negative training samples when learning the traditional correlation filters,and the error accumulation caused by linear updating during tracking,with the application of spatial-temporal context information of the target,this dissertation studying the adaptive suppression of background critical distractors,and the coupling of multiple dynamic target templates,by which proposing a weight adapted distractor-aware template-coupled correlation filter joint learning model,improving the accuracy of the correlation filter for target recognition and the ability to suppress the background key distractors,and also achieving non-linear updating for the learned filters.Based on the filter set learned by DATC-CF,this dissertation proposes a multi-model tracking algorithm DATC＿MM with the optimal filter adaptive selection strategy.During tracking,the model updating strategy based on high confidence criterion and the target redetection strategy based on a complementary color detector further improve the robustness of the proposed tracking algorithm.The experimental results based on a large number of datasets,such as OTB2013,OTB2015,Temp-Color,DTB70 and NFS30,demonstrate the adaptability of the proposed tracking algorithm DATC＿MM to complex scenes,and the comparison results with current state-of-the-art tracking algorithms verify the superiority of DATC＿MM further.(4)Considering the boundary effect caused by Fourier transform in the traditional correlation filtering tracking when learning the filters,as well as the poor representation efficiency and information redundancy of pre-trained CNN features,this dissertation propose a robust tracking algorithm based on spatial-temporal feature perception and convolution residual fusion.In this algorithm,based on the regression loss and ranking loss of the training samples,a spatial-temporal feature perception learning network is constructed to refine the off-line pre-trained CNN features.With the convolution residual learning network and the learned spatial-temporal CNN features,a spatial-temporal convolutional residual learning network is obtained,based on which a robust tracking algorithm STCRL is implemented.STCRL can adapt to the changes of target appearance in real time,and train the correlation filter in the time domain,avoiding the boundary effect problem of the traditional correlation filters.Experimental results on several datasets demonstrate the effectiveness and advancement of the proposed algorithm.