Real-time Visual Object Tracking Exploiting Superpixel And Correlation Filter

Visual object tracking is one of the most active topic in computer vision and has been widely used in real-world applications,such as human-computer interaction and intelligent monitoring.In the area of visual tracking,researchers have proposed many state-of-the-art algorithms,but due to various challenging aspects,such as stability-plasticity dilemma in target representation,stochasticity in background,and balance problem between tracking precision and speed,the existing algorithms cannot well realize real-time robust visual tracking.Therefore,visual object tracking remains a challenging topic for further research.To comprehensively address various challenging problems in visual tracking and to realize accurately real-time visual tracking,this thesis introduces a novel visual tracking framework and proposes real-time visual tracking based on superpixel and correlation filter.Compared with existing visual tracking algorithms,the contributions of this thesis are as follows:First,we propose a novel visual object tracking framework,i.e.a joint prediction-detectioncorrection tracking framework.The predictor is used to estimate motion offset and scale variation of target,and these predicted parameters are passed to the detector;Combining with those predicted parameters,the detector is used to determine target's location and size;The corrector is used to detect target location with classifiers which are trained online,and the corrected result is refined to further determine the target location obtained in the detector.Via this tracking framework,we can realize robust long-term tracking and well address problems such as target occlusion and out-of-view problems.Second,we introduce a method to predict target motion and scale variation based on superpixel analysis and optical flow.Superpixel analysis is used for target appearance model reconstruction,and extended superpixel centers are used for optical flow calculation to estimate target motion offset and scale variation.This method can effective predict the target's motion trend and scale variation,and these predicted parameters can be used for subsequent accurate tracking.Third,we improve kernelized correlation tracking based on superpixel predicted parameters.Since kernelized correlation tracking cannot well address wide-range motion and scale variation,we integrate predicted parameters into kernelized correlation filter,construct position correlation filter and scale correlation filter,and introduce an adaptive learning rate into target model update scheme,so that we improve the performance of kernelized correlation tracking and realize robust real-time tracking.Fourth,we devise support vector machines to construct a correlator exploiting dense sampling examples and reliable history examples.According to response values of correlation filter,we train dual support vector machines online for fast detection through passive-aggressive learning and re-training,and then combine corrected results using refine mechanism for accurate target location.Extensive experiments are conducted on Object Tracking Benchmark(OTB-2013),TempleColor128 and TB-100.These comprehensive experiments include experiments of overall quantitative analysis,attribute-based quantitative analysis,overall qualitative analysis,realtime performance analysis,parameters effectiveness analysis,module effectiveness analysis.These experimental results demonstrate that the proposed visual tracking algorithm,i.e.PDCTs,can achieve state-of-the-art performance and realize high-precision and high-success-rate visual object tracking,while maintains good real-time performance.Besides,the experimental results prove the effectiveness of each module and the rationality of each parameter in the proposed algorithm,and also indicate that the proposed algorithm has the reliability to full address various challenging aspects.In addition,this thesis also explores the application of the proposed visual tracking algorithm in an in-air-writing system.Since single finger writing system cannot well mimic actual writing situation,we propose the pinched fingertips gesture for the in-air-writing system.We utilize the proposed PDCTs for pinched fingertip tracking,extract the trajectory to realize vision-based inair-writing system and to demonstrate the practicability of the proposed visual tracking algorithm.