Font Size: a A A

Object Tracking Based On Multi-feature Fusion And Particle Filter

Posted on:2015-06-11Degree:MasterType:Thesis
Country:ChinaCandidate:D L DaiFull Text:PDF
GTID:2298330422471509Subject:Control engineering
Abstract/Summary:PDF Full Text Request
With the growing popularity of imaging device and the rapid development ofimage processing theory in recent years, object tracking has become a core technologyin the field of computer vision. Object tracking technology is widely used inmonitoring for security, intelligent transportation, guidance for weapon, medicalauxiliary and so on. This paper introduces some typical tracking algorithms, such asmean shift belonging to optimal matching based mean, kalman filter and particle filterbelonging to motion prediction based mean, and then points out that particle filter isthe best theory framework for object tracking. We study on the tracking algorithmbased on adaptive fusion of multiple features in the framework of particle filter, andfocus on three key issues–acquisition of visual features, fusion of multiple featuresand robustness of particle filter.Selection and acquisition of visual feature usually have an important effect ontracking performance, as features express some intrinsic information of object. Thispaper introduces how to construct histograms for color, texture and edge featurerespectively. Expressly, in order to obtain fuzzy edge in the image better, we study onthe fuzzy theory based edge detection. Aiming at the shortcoming of general fuzzyinference based edge detection in inference strategy and defuzzification, we have madeimprovements and then proposed an edge feature acquisition algorithm based onimportance weighting and fuzzy inference step-by-step, which increases the accuracyof edge feature expression greatly.However, any visual feature can’t express object information perfectly, so singlefeature based tracking algorithm is difficult to achieve good tracking result whenimage scene is complex. Generally, for various visual features, taking advantage oftheir inherent complementarity and fusing them with some strategy is a good way toenhance the accuracy and robustness of object tracking. Relying on probabilityfunction for system observation in the framework of particle filter, this paper hasproposed an adaptive fusion strategy for multiple features based on log-likelihood ratio.Its superiority is not only considering the similarity between candidate template andobject template, but also considering the distinction between candidate template andbackground template.In addition, when filtering process is heavily influenced by external disturbance, some particles far away from real status of the object usually own large weights in theparticle filter based tracking algorithm. Under this circumstance, particle resamplingmay cause filter divergence even tracking error. To address this problem, this paperemploys kernel-based particle filter to carry out object tracking. Before particleresampling, it calculates the distance between each particle and current state of object,and then adjusts particle weights with kernel function. By this way, the convergence offilter and the robustness of tracking can be greatly improved.
Keywords/Search Tags:Object Tracking, Particle Filter, Multi-Feature Fusion, Edge Acquisition
PDF Full Text Request
Related items