Research On Tracking Algorithm For Visual Targets Based On Random Finite Set

Visual tracking can be defined as the problem of effectively tracking and estimating stateparameters of targets of interests like position, velocity, size etc in different frames of a video,which is an important basis of fulfilling higher levels of tasks such as action recognition andanalysis, as well as pose estimation etc. Visual tracking is an active topic in computer visionfield, which enables several important applications such as visual surveillance, militaryguidance, human-operator interaction, visual navigation of robots, and traffic management etc.Despite the wide applications, object tracking can be very difficult, due to reasons such ascomplex tracking environment, dynamic target regions, and a varying number of targets.Recently, the random finite set (RFS) based multi-target Bayes filters and theirimplementations have been a promising alternative to the traditional association-basedmethods. The RFS based multi-target Bayes filter is able to directly track multiple targetsavoiding data association techniques. In the finite set statistics (FISST), the RFS model for thetime evolution of the multi-target state incorporates target motion, birth and death, thereforethe RFS based multi-target Bayes filter is suitable to be employed to track a variable numberof targets. However, since the conventional RFS based multi-target Bayes filters are aimed atdealing with tracking point objects, complex tracking scenes and targets with scales posegreat challenges on RFS based visual tracking algorithms.To improve the robustness of RFS based multi-target tracking algorithms and furtherprovide supports for research and applications, the dissertation proposes some enhancedvisual multi-target tracking algorithms based on RFS. These proposed algorithms achieve thetracking performance improvement via three aspects, namely, adding trajectory recognitionability to RFS based filters, building better discriminating and reliable likelihood model, anddesigning more robust and accurate state extraction algorithms. By incorporating estimate-track association, adaptive multifeature fusion, adaptive update of models and kerneldensity estimation techniques, we propose the Gaussian mixture probability hypothesisdensity (GM-PHD) based visual tracking algorithm with trajectory recognition, adaptivemultifeature fusion based PHD visual tracking algorithm, particle PHD visual trackingalgorithm with color feature and Gaussian mixture model based clustering, and particle PHDbased visual tracking algorithm with kernel density approximation, which are able toefficiently track a varying number of targets in complex scenes. The main contributions ofthis dissertation are summarized as follows:1. Based on Gaussian mixture PHD filter, a robust multi-target visual tracking algorithmwith trajectory recognition is proposed. The proposed algorithm uses position estimatesgenerated by detectors as measurements. To increase the robustness of the trackingmethod, we improve the background subtraction based detector for more accurate targetposition observations via updating the background model only using the backgroundpixels in the current frame of the video sequence. Since the state estimate RFS of targetsfrom the PHD filter can not provide any information on the identity of targets, we proposea track recognition method based on auction algorithm, which can be taken as a“estimate-track” association method. Incorporating the proposed “estimate-track”association method into the PHD filter, the PHD filter is extended to be a tracking methodwith the ability of recognizing targets’ tracks, and the robustness of detection algorithmbased PHD tracker is improved for more clutters are removed from the state estimates ofPHD filter.2. To overcome the shortcomings of lower degree of descriptiveness and degradeddistinctiveness of the target model only using single visual cue in complex environments,a tracking algorithm based on adaptive multifeature fusion is proposed. In complex scenes,a single cue based visual tracking method may lead to failures for the likelihood can noteffectively discriminate targets from images with background and clutter. The colorfeature (global feature) and scale invariant feature (local feature) are utilized to representthe target, and the two features associated likelihoods are designed. The reliabilitymeasure based on spatial uncertainty is applied to weight the reliability of each featureand integrate likelihoods of features in a weighted sum way with their weights on-lineadjusted according to reliability scores of the features. Incorporating the likelihoodfunction with adaptive cue fusion mechanism into the particle PHD filter enhances the robustness of the tracking algorithm. The proposed algorithm not only is able toefficiently track a varying number of targets, but can robustly track targets when varyingscales of targets, out-of-plane target rotation, occlusion and background region withsimilar appearance to targets occur.3. Particle PHD based visual tracker with color feature and Gaussian mixture modelclustering is proposed. To avoid traditional PHD based visual trackers’ dependence on theadopted detection methods, we combine color feature with PHD filter in a unifyingframework by designing likelihood using color histogram, which increases the flexibilityof PHD based visual tracker and distinctiveness of the target model with respect to thebackground and clutter. Aim to deal with the unreliability of clustering techniques forextracting state estimates in the particle PHD filter, a robust state extraction algorithmbased on Gaussian mixture model (GMM) clustering is proposed. In the GMM clusteringbased state extraction algorithm, considering the characteristics of visual tracking, wedesign a Gaussian component management strategy for initializing model update method,namely expectation maximization, which improves reliability and accuracy of theextracted state estimates. We combine color likelihood and robust state extraction methodwith the particle PHD filter, thus improving the robustness and flexibility of the particlePHD based visual tracker.4. To increase the robustness of the particle PHD filter based visual tracker, particle PHDbased visual tracking algorithm with kernel density approximation is proposed. Due to themultimodality distribution of the multi-target states, as well as occlusion and clutter incomplex scenes, the distribution of the resampled particles describing the posterior PHDis much more complicated. Considering the complex distribution of particles, we modelextracting state estimates from the resampled particles as an accurate density estimationproblem. Generally, parametric density estimation algorithms like Gaussian mixturemodel based method require prior knowledge on the form and the component number ofthe underlying mixture density function. In addition, the convergence and accuracy of theparametric density estimation algorithms depend on initialization. But kernel densityestimation algorithm, as a non-parametric method, need not to know any information onthe underlying density function, and can effectively represent any arbitrary densityfunction since the estimated density is mainly dependent on the structure of data itself.Due to the accuracy and flexibility of kernel density estimation, we apply it to design a robust state extraction algorithm based on incremental kernel density approximation(IKDA) for particle PHD filter. By incorporating the IKDA based state extraction method,the particle PHD based visual tracking algorithm is improved on its robustness.