基于在线多示例学习的协同训练目标跟踪算法  

Co-Training Object Tracking with Online Multiple Instance Learning

作　　者：李飞[1] 王从庆[1] 周鑫[1] 周大可[1] 

机构地区：[1]南京航空航天大学自动化学院,南京210016

出　　处：《吉林大学学报（信息科学版）》2015年第2期201-207,共7页Journal of Jilin University（Information Science Edition）

基　　金：国家自然科学基金资助项目(61172135)

摘　　要：为解决把多示例学习应用到目标跟踪算法而导致的误差积累问题,结合协同训练方法,提出一种新的目标跟踪算法。该算法利用协同训练克服分类器自训练带来的误差积累,同时在线多示例学习提高了跟踪效果的鲁棒性。将跟踪结果中心与理想目标位置中心的误差作为评价标准,在标准视频序列上将跟踪结果与半监督学习跟踪算法和传统多示例学习跟踪算法进行对比。实验结果表明,该方法在背景光照变化、目标旋转等复杂条件下,可很好地跟踪目标,具有较好的鲁棒性。To solve the problem that multiple instance leafing can cause error accumulation in object tracking algorithm, a new object tracking algorithm was proposed based on co-training and online multiple instance learning. This algorithm uses co-training to overcome errors accumulation caused by self-training and improve the robustness of tracking performance based on online multiple instance learning. The center error between tracking results and ideal object location was used as evaluation criteria. The semi-supervised learning tracking algorithm and traditional multiple instance learning tracking algorithm were simulated on the video frames which come from standard video library. The tracking results show that the algorithm performance is more superior, and the center error curves further demonstrate the experimental results. Experimental results show that the proposed approach can track the target well in the complex conditions, and it has a better robustness.

关 键 词：多示例学习 协同训练 目标跟踪 在线学习 目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]