在线加权多示例学习实时目标跟踪  

Real-time object tracking via online weighted multiple instance learning

作　　者：陈东成[1,2,3] 朱明[1,2] 高文[2] 孙宏海[2] 杨文波[1,2,3] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所中国科学院航空光学成像与测量重点实验室,吉林长春130033 [2]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [3]中国科学院大学,北京100039

出　　处：《光学精密工程》2014年第6期1661-1667,共7页Optics and Precision Engineering

基　　金：中国科学院航空光学成像与测量重点实验室开放基金面上项目(No.Y2HC1SR121)

摘　　要：由于原始多示例学习(MIL)跟踪的分类效果和实时性较差,提出了一种加权在线多示例学习跟踪算法。首先,根据所选定目标位置分别采集目标和背景样本集,通过对所采集样本集特征的在线学习生成弱分类器集;然后,用计算样本集对数似然函数的最大值的方法从弱分类器集中选择K个最优的弱分类器,给每个弱分类器赋不同的权值,生成一个强分类器;最后,在新的一帧中抽取目标和背景样本,用生成的强分类器对待分类的目标和背景进行分类;分类结果映射成概率值,概率最大样本的位置就是所要跟踪目标的位置。对不同视频序列的测试结果表明,该跟踪算法的跟踪正确率达93%,目标大小为43pixel×36pixel时处理帧率约为25frame/s。与原始多示例学习跟踪算法相比,本算法的实时性提高了67%。Abstract： A weighted Multiple Instance Learning（MIL） tracking method was proposed to improve the precision and real-time quality of online MIL tracking algorithm. First, target samples and back-ground samples around a selected target were collected. Weak classifiers were generated by online learning the features of collected samples. In order to get K best weak classifiers, the maximum of samplest log-likelihood was calculated. Every weak classifier was weighted differently and K weak classifiers were combined into a strong classifier. Finally, new unclassified samples were picked from the newly formed frame. The obtained strong classifier was used to separate the target and background. The classifying results were mapped into probabilities and the location of the sample with the largest probability was the target location wanted. Experiments on variant videos show that the accurate rate of the proposed algorithm is 93% and the average frame rate is 25 frame/s when the object size is 43 pixel× 36 pixel. Compared with the original MILtracking algorithm, the real-time quality of proposed method increases by 67%.

关 键 词：多示例学习 目标跟踪 分类器 权值 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]