Research On Novel Method Of Visual Tracking And Its Applications

Visual tracking is an important topic in the field of computer vision，it has a wide rangeof application in military guidance, visual surveillance, visual navigation of robots,human-computer interaction and medical diagnose, etc. The goal of visual tracking is toenable the computer to imitate the motion sensibility of human vision, perceive the movingtarget in a video, and provide and important data source for visual analysis and understanding.In recent years, with the rapid growth of the computer techniques and sensor techniques,visual tracking has attracted many researchers’ attention, has become a very popular researchproblem. A lot of new theories and excellent visual object tracking methods have beenproposed. However, Visual target tracking often become very difficult due to many factorssuch as complex background, illumination variation, object rotation, occlusion and randommotion. Many problems and difficulties in theory research and in applications are stillunsolved. It’s great challenge to build a robust, precise, stable and practical visual trackingalgorithm, which is both theoretically and practically valuable.On the basis of traditional visual tracking methods, the thesis aims to two mainalgorithms: particle filter and mean shift. It combines academic frontiers and presents newidea and method to improve the accuracy and robustness of object tracking. The maincontents and contributions of this dissertation are summarized as follows:Firstly, four new visual tracking algorithms with a framework using particle filter areproposed.(1) In order to improve the quality of particle sampling and the accuracy of visualtracking, a ball particle filter algorithm for visual tracking is proposed. Ball sampling modeintroduced can guarantee the valid particles in state-space. Compared to the conventionalparticle filter, the proposed method used much fewer particles to ameliorate the diversity ofdistribution, and overcame the degeneration problem effectively. By iterative motion of ball,particles are moved towards regions where they have larger values of posterior densityfunction. Ball particle filter which does not depend on state-mode can track maneuver objectwhich movement is irregular. The proposed method can improve the efficiency of particlesand achieves preferable precision of tracking.(2) A particle filter for object tracking based on multi-region sampling is proposed tosolve the problems of degeneracy phenomenon and particle impoverishment introduced bytraditional particle filter algorithm. The proposed method uses some overlapping sub-regionsto divide the target model, and each sub-region corresponds to a sampling windows. The truestate of target can be estimated by the confidence of each sub-region. The complementary andstage uniqueness of sub-region can guarantee the validity of particles and the quality ofstate-space. Thereby, the accuracy of object tracking is improved. The proposed methodrelieves effectively the sample degradation and poverty problems, improves the accuracy ofvisual tracking by effective particles.(3) A new anti-occlusion method for object tracking is presented to solve the problemthat traditional visual tracking algorithm often deviates or loses the targets under occlusions. The high likelihood areas generated by the SSD residual can adjust the range and quantity ofparticle sampling in the state-space. The sampling method can cover various possibilities ofobject state and improve the quality of the state-space exploration in the diffusion process ofthe particle filter. The object state of forecast and estimation fused by noise covariance canachieve reliable tracking performance under occlusion and gain the optimal location of object.The adaptive quantity of particle sampling not only can improve the precision, but also canreduce the computational load in a certain extent effectively. The proposed method has strongrobust and error-tolerance to occlusion of tracking objects, and has good performances undercomplex background.(4) In order to avoid the poor robustness based on most of present observation models insmall sample space, a particle filter algorithm for visual tracking based on partial featurecombination is proposed. Partial features can represent the detail of target template effectively,and can alleviate the affection of object deformation, illumination change and partialocclusion in feature matching. The proposed method employs the idea of mixture of Gaussianand uses multiple modes to represent valid partial observation information. The strategy offusion is more precise and reliable, thus can overcome the degeneracy problem by newmeasurement and improve the efficiency of object tracking. The small sample space canreduce quantity of particle and computational load in a certain extent. The proposed method ismore effective than tracking algorithm with single feature or common multi-features fusion.Secondly, two improved algorithms are proposed to improve the performance of meanshift.(1) Considering the issue of template matching within the Mean Shift framework, thispaper proposes a concept of feature contribution. It can effectively reduce the influence ofbackground feature and noise, make importance feature play a key role. In addition, binarydistribution of structure introduced can effectively reduce the error of statistical features byspecial information and improve the tracking accuracy and robustness in a certain extent.(2) A new tracking algorithm based on double-ring Mean Shift is proposed in this paperto solve the deficiency of target representation, template similarity measure and fixedkernel-bandwidth in traditional Mean Shift tracking algorithm. The feature extraction modelbased on universal elliptical region is used in this algorithm to reduce the influence ofbackground feature and improve the quality of target model effectively. Double-ringdescriptor is presented to emphasize the importance of target feature and improve the peakmodality of matching function. The proposed method can update the bandwidth ofkernel-function adaptively by the relationship of double-ring. The proposed tracking approachis robust and invariant to scale, pose and partial occlusions.