基于连续最小能量的多目标跟踪  

Continuous energy minimization based multi-target tracking

作　　者：孙娴[1] 朱松豪[1] 金栋梁[1] 

机构地区：[1]南京邮电大学自动化学院,江苏南京210023

出　　处：《南京邮电大学学报（自然科学版）》2016年第1期111-118,共8页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(31200496);国家博士后基金(2014M550297);江苏省博士后基金(1302087B)资助项目

摘　　要：文中针对多目标跟踪问题,提出基于连续最小能量的多目标跟踪方法。该算法首先利用观测模型、表观模型、运动模型、互斥模型、轨迹维持模型及轨迹修正模型构建一个目标函数;然后利用梯度下降法对构建的目标函数求解,以得到各时刻跟踪目标的近似最小能量及对应的多目标个数和状态;最后采用基于最小能量的智能探测方法,得到平滑、连续的跟踪轨迹。PETS2009/2010benchmark和TUD-Stadtmitte视频序列库实验结果表明,文中提出的方法能够在存在杂波、虚警和漏检的复杂场景中实现多目标的正确关联,得到稳定、持续的跟踪轨迹。A method to deal with the issue of multi-target tracking by considering continuous energy minimization is proposed. Firstly,observation,appearance,dynamic,mutual exclusion,trajectory persistence,and trajectory regulation models are integrated into an objective function. Then,the gradient descent method is utilized to solve the constructed objective function to obtain the approximate minimum energy at every moment,and obtain the number and status of multi-targets. Finally,the intelligent extrapolation method based on the continuous energy minimization is utilized to achieve the final smoother and more continuous tracking trajectories. Experimental results on PETS2009 /2010 and TUD-Stadtmitte video database demonstrate that the effectiveness and the efficiency of the proposed MAP-HOR scheme.

关 键 词：连续最小能量 梯度下降法 探测法 误警 轨迹 

分 类 号：TP391.412[自动化与计算机技术—计算机应用技术]