Multi-target Tracking By Hierarchical Association Method

Scholars pay attention to the hot topic-Target Tracking Technology in the field of computervision and pattern recognition research and many of them think it had a broad application prospects.The Multi-Target Video Tracking has been extensively studied in recent years, and many newtracking algorithms has been put forward, but stable and accurate tracking results is still verydifficult to obtain. From the point of multi-target tracking and multi-target association of videodetection, this paper puts a conducted in-depth research on multi-target tracking technology andproposed practical and effective solutions to the overlap of target, identity switches, missed targetsand false alarms.This paper has taken a research on Multi-Target Tracking by Continuous Energy Minimization.Research method is to predict the state of target’s motion through target detection. This methodintegrates the multi-target observation model, dynamic model, mutual exclusion model, targetpersistence model and regulation model within the same objective function. It solves the objectivefunction by standard conjugate gradient to obtain every moment of multi-target approximateminimum energy, and thereby obtained a target number and status. As for target occlusion, the paperproposes Target Occlusion of Analytical Global Occlusion Model. Through an analysis on overallvisibility by each target, the paper find one in the continuously differentiability in closed form thatis in line with piecewise differentiable closed interval indicator Gaussian function. Finally, the paperobtains the target detection information by finding the derivative of the indicator function. Thismethod also takes into account the state of motion and apparent model, which is used to avoiddetection target false alarm.For the issue of multi-target tracking trajectories generation, this paper proposes a hierarchicalassociation method based on apparent model. On the basis of detection response and initialtrajectory extraction, the method uses OLDAMs and AdBoost algorithms to realize the initialassociation goals. Because the apparent discrimination online tracking model trajectories are toofragmented and discontinuous, this paper uses short and reliable tracking fragments to realizesecondary association. As the second association is not suitable for a long time target trackingprocess, this paper gets the final smoother intelligent continuous tracking trajectory detectionmethod by extrapolation.In addition, the paper also puts the method into PETS2009/2010benchmark and TUD-Stadtmitte video sequence database to prove that the method could achieve multipleobjectives correct association in the presence of clutter, false alarm, missed in complex scenes,thereby obtaining stable, continuous tracking tracks.