基于可变形部件模型的内河多船舶跟踪算法研究  

Multiple Ship Tracking in Inland Waterway via Deformable Part Model

作　　者：朱琳[1] 郭建明[1] 刘清[1] 李静[1] 

机构地区：[1]武汉理工大学自动化学院,武汉430070

出　　处：《计算机科学》2015年第3期311-315,共5页Computer Science

基　　金：国家自然科学基金项目(51279152);武汉理工大学自主创新研究基金项目(145211005)资助

摘　　要：内河航运事业的不断发展对电子巡航闭路电视(CCTV)监控系统的智能化水平提出了越来越高的要求。针对目前CCTV监控系统智能化水平较低、人工参与量较大的局限,提出了一种鲁棒的基于可变形部件模型的内河多船舶跟踪算法。该算法将每条运动目标船舶视为一个部分,通过最小生成树模型来建立各个部分之间的联系,并基于可变形部件模型的原理最终实现同时对多条目标船舶的鲁棒跟踪。为了得到精确的目标船舶外观模型,首先对目标区域进行梯度方向直方图(HOG)特征提取,然后利用模糊支持向量机(SVM)算法进行训练得到每条目标船舶的参数化的外观模型。其中,由于模糊SVM中模糊度的引入,对不同的输入训练样本赋予不同的重要性,因此将获得比线性SVM算法更加精确的目标外观模型。结构化的学习方法保证了在目标运动过程中该算法能即时更新目标间的空间相互关系参数,实现鲁棒的跟踪效果。实验结果表明,提出的算法适用于在内河环境下鲁棒、有效的多目标船舶跟踪。The closed-circuit television（CCTV）surveillance system is developing rapidly in recent years.But the intelligence level is relatively low.In this paper,a robust multiple ship tracking algorithm was proposed based on the deformable part model.The proposed algorithm treats every ship as a part.By incorporating the spatial constraints,the interrelations model between ships with the minimum spanning tree model can be effectively built.Then the robust multiple ship tracking is accomplished based on the deformable part model.Moreover,aiming at obtaining an accurate parameterized appearance model of ships,the HOG features combined with the fuzzy SVM is adopted to train those object regions.Especially,because of the ambiguity in the fuzzy SVM,every training samples are given different importance so as to obtain a more accurate appearance model.At the same time,structured learning can guarantee to update the interrelation parameters on time when ships move.Experimental results demonstrate that our proposed algorithm is suitable for inland waterway and can accomplish robust and effective multiple ship tracking.

关 键 词：内河 闭路电视监控系统 多船舶跟踪 模糊SVM 可变形部件模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]