机构地区: 通信与信息工程学院
出 处: 《西安邮电大学学报》 2014年第2期43-47,共5页
摘 要: 为提高运动目标跟踪算法的鲁棒性,提出一种基于多示例学习(MIL)框架的跟踪算法。该算法利用类Haar特征构建若干弱分类器,然后级联为多示例学习强分类器,根据目标在视频前一帧中的位置,依据最大熵原理,在当前帧中找出目标可能出现的范围,并利用该强分类器确定其最有可能出现的位置,作为跟踪结果,并且将该位置不同邻域内的图像分别作为正包和负包去更新多示例学习强分类器。实验结果表明,该算法对于运动目标外观有显著变化的情况具有较好的鲁棒性和实时性。 In this paper,an object tracking algorithm with multiple instance learning is proposed. This algorithm uses Haar-like features to build many weak classifiers,and combines some of them into an additive strong classifier.According to the position of the object in the former frame,the position of the object in the current frame is ascertained with the strong classifier.Different image patches are cropped as positive and negative bags to update the online strong classifier in its neighborhood.The experimental results show that this algorithm has good robustness and real-time performance when the appearance of object undergoes some significant changes.
领 域: [自动化与计算机技术] [自动化与计算机技术]