基于约束知识的IP-MCMC-PF目标跟踪方法研究  

Research on target tracking based on constraint knowledge IP-MCMC-PF

作　　者：梁启香 汪荣贵[1] 张冬梅[1] 李想[1] 秦飞[1] 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机工程与应用》2016年第1期120-126,265,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61075032);中央高校基本科研业务费专项资金(No.2012HGCX0001)

摘　　要：针对粒子滤波算法在有干扰的目标跟踪中可能出现的粒子多样性减少和精度下降等问题,研究并实现了一种新的基于约束知识的IP-MCMC-PF目标跟踪方法。该方法首先通过约束知识提高粒子预测的准确性,并通过多链并行的IP-MCMC方法提高粒子的多样性,有效地解决粒子退化问题;在此基础上通过PN学习算法在线更新抽样粒子的抽样分布和检测器的训练样本,实现目标跟踪算法的在线学习,有效提高了复杂背景下目标跟踪的准确度和自适应性。实验结果表明,该方法在遮挡、形变、光照变化等多种干扰的情形下都具有很好的跟踪效果。Particle filter algorithm for target tracking with interference may occur in the problem such as lower particle diversity and reduced precision. In allusion to this instance, a novel tracking method based on constraint knowledge is proposed. This method improves the precision of the particle prediction using constraint knowledge. The problem of the particle degeneration is effectively solved by improving the particle diversity with the parallel IP-MCMC method. On this basis,the proposed method realizes the online study algorithm using PN learning, which is used to update the sample distribution of particles and the training samples of the detector. The accuracy and adaptability of the tracking method under complex background is effectively improved. Experimental results show that the proposed method has good effect under the situation of various interference（e.g., shade, deformation, illumination change）.

关 键 词：粒子滤波 约束知识 IP-MCMC抽样 PN学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]