Conventional tracking methods describe the target with a bounding box .As the bounding box is likely to contain some background regions and will degrade the tracking performance ,a superpixel tracking method via online multiple instance learning is proposed.In training stage,input frame is segmented into superpixels,which are divided into several instance bags with clear labels according to their location .The tracking is thus converted into a multiple instance learning problem .Then,online multiple instance learning is implemented with the algorithm.The maximum of instance bags’log-likelihood function is calculated to get K best weak classifi-ers,which are combined into a strong classifier .In detection stage ,a confidence map is generated by the strong classifier in the subsequent frame .Finally,the state of the tracking target is estimated with the confidence map in particle filter framework.The proposed method runs at a rate of 15 frames per second on a laptop.Extensive experimental results on challenging sequences show that the proposed method performs well in terms of robustness and accuracy,especially for the target under complex background ,moving at high-speed or is occluded .Compared with the original superpixel tracking,the typical values of precision and success rate of the proposed method are increased by 21%and 26%,reaching 91%and 90%,respectively.%采用矩形框表示目标会引入背景干扰，导致跟踪性能下降，故利用多实例学习的特点对背景干扰建模，提出了一种采用在线多实例学习的超像素跟踪算法。在训练阶段，以超像素作为实例，根据位置将这些超像素分为具有明确标签的多个实例包，进而将跟踪转换为多实例学习问题。然后，在所提算法中实现了在线多实例学习，通过求实例包的似然函数最大化，从弱分类器池中选择K个最优的弱分类器组合为强分类器，在下一帧的检测阶段，利用学习的强分类器生成目标置信图。最后，采用粒子滤波方法从置信图中估计目标状态，在2．6 GHz主频的笔记本电脑上，所提算法的跟踪速率可达15 frame／s。在多个视频序列上的对比实验表明，该算法对复杂背景、目标高速运动、遮挡等具有更好的鲁棒性和精度，且跟踪精度和成功率的典型值分别达到了91％和90％，比原始超像素跟踪算法分别高出了21％和26％。
展开▼
