| Recent advances in computer vision theory have boosted its usage in different areas,such as military defense,secure surveillance,robotics and so on.Multi-target visual tracking,as one of the most fundamental techniques in image processing,is of critical importance for behavior understanding,scene analysis and other high-level applications.However,most research on multi-target visual tracking only focus a particular issue in a specified tracking scenario,and only few work concern about the tracking framework.With the development of the Random Finite Set(RFS)based Bayesian framework,different filters have been proposed for multi-target tracking in the radar and sonar system and proved to be effective and efficient.Therefore,the author takes the advantage of RFS based Bayesian filter and extends it to the visual tracking area.This thesis studies the RFS based multi-target visual tracking and trajectory fusion,in order to propose a general online tracking framework as a solid foundation in the visual tracking community.The main contents of this thesis are given as followings in details:Firstly,we consider the δ-GLMB and LMB filters for point process multi-target tracking with trajectory maintaining,which generate a large number of association hypotheses to perform track maintenance and the computational cost is barely unacceptable for real-time online multi-target tracking.So,this thesis proposes an efficient implementation named as the Labeled Probability Hypothesis Density(LPHD)filter based on the δ-GLMB and LMB filters.Under the assumption that the dynamic model and measurement model are(or nearly)both linear and Gaussian,the Gaussian Mixture implementation of the LPHD filter is given in details,where a single Gaussian component to approximate the each target posterior in order to speed-up the update process by reduce the computational cost caused by many posterior association hypotheses.Simulation results validate the effectiveness and efficiency of the LPHD filter compared with the δ-GLMB and LMB filters.Secondly,this thesis applies the LPHD filter to the visual tracking problem as a general framework and proposes the LPHD visual tracker.By comparing and analyzing different target representations of visual targets,the rectangle descriptor is used for target location and shape,and the feature histogram is adopted to describe target’s visual appearance.An adaptive method is used to handle target trajectory initialization by using the measurements at each frame,which is independent from the prediction process.As for multi-target state prediction,the dynamic model and target survival probability are built upon target location and size.In the update process,the measurement model contains both motion similarity and feature similarity,and the detection probability is established by modeling the mutual occlusion among targets.Experiments compared with the state-of-the-art trackers have shown that the LPHD visual tracker has achieved equivalent tracking precision and outstanding efficiency.In addition,the parameter study is carried out to show the influence on the tracking performance of the LPHD filer.Thirdly,the LPHD visual tracker is studied to deal with tracking-by-detection process by blending in online-trained detector,in order to alleviate the impact that the offline-trained detector has on the tracking performance.With the Kernelized Correlation Filter(KCF)being the online detector,the LPHD-KCF visual tracker is proposed to perform online detection and tracking simultaneously.Each existing target is associated with a unique KCF,and the associated KCF generates pseudo-measurement to update the target state and appearance model.The LPHD update process utilizes the measurements obtained by off-line trained detector to determine which targets are new born targets that need to be initialized its own KCF.Besides,the validation parameter is proposed based on the KCF response in order to remove KCFs that are already drifting.Experiment results have shown that the LPHD-KCF visual tracker can gain better trajectory association precision when the tracking scenario is difficult for off-line trained detector.Nevertheless,the computational cost is huge due to the online-training process,especially for dense targets.At last,this thesis further studies the multi-sensor fusion schemes of the LPHD filter to deal with the multi-view multi-target trajectory fusion problem by considering each calibrated cameras as position sensors.As for the centralized information fusion,this thesis proposes seq-LPHD,para-LPHD and rand-LPHD according to the classical sequential update,parallel update and random update respectively.With regard to the distributed information fusion,the Kullback-Leibler Average(KLA)divergence is studied to guarantee each sensing node converges to global consensus after several iterations of information exchange and local fusion.Thus,the KLA-LPHD is proposed based on the information consensus to handle distributed fusion problem.Simulation results have validated the effectiveness of all these four multi-sensor LPHD fusion methods.Further experiments are carried out on multi-view pedestrian tracking dataset which indicate that the multi-sensor LPHD filter is capable of multi-view target trajectory fusion with proper detector. |
