融合SIFT特征和HMM的运动目标识别与跟踪算法  

Moving Target Recognition and Tracking Algorithm Based on Fusing SIFT Features and the Hidden Markov Model

作　　者：顾苏杭[1,2] 戎海龙 马正华 

机构地区：[1]常州大学信息科学与工程学院,江苏常州213164 [2]常州轻工职业技术学院信息系,江苏常州213164

出　　处：《科技通报》2017年第11期210-215,共6页Bulletin of Science and Technology

基　　金：2014年度到2016年度国家自然科学基金项目(51307010);2014年度到2017年度江苏省自然科学基金项目(BK20140265);2016年度到2018年度常州市科技计划资助项目(CJ20160010)

摘　　要：针对动态场景下,由于光照突变、目标旋转以及遮挡等因素容易导致运动目标跟踪丢失,本文提出融合SIFT特征和隐马尔科夫模型算法。将运动目标的SIFT特征作为隐马尔科夫模型训练样本,经训练得到特征最优化的模型参数;通过设定阈值,将模型输出较大计算概率特征样本的集合作为最终目标识别结果,不仅避免了SIFT算法遍历式处理图像像素点带来的大量计算,而且该样本集能够精确反应出目标区域位置信息,从而取代了SIFT算法图像间繁琐的匹配过程,提高目标跟踪的可靠性和稳定性。实验结果表明,目标平均识别率在90%以上,跟踪效果稳定、可靠,具有较好的实时性和鲁棒性。In order to solve the moving target tracking lost under dynamic scenes due to factors such as illumination, target rotation and occlusion, this paper proposes a new algorithm that fuses SIFT features and the Hidden Markov Model algorithm(HMM). Taking SIFT features of moving target as training samples of HMM, then getting the optimal model parameters by model training; And through setting threshold value, taking collection of larger calculation probability features that HMM outputs as eventually target recognition results. The algorithm can ensure not only avoiding a large number of computing that brought by SIFT algorithm iterated through pixel to process image, but also the sample set can accurately reflect the regional location information of target which can replace the tedious matching process between continuous images of SIFT algorithm. And it improves the reliability and stability of target tracking. Experimental results demonstrate that the target average recognition rate above 90%,tracking performance is stable and reliable, and having better real-time performance and robustness with using the new algorithm.

关 键 词：动态场景 目标旋转 隐马尔科夫模型 识别与跟踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]