摘要
由于目标小、可区分性差,无人机对地目标跟踪较传统视频目标跟踪更容易丢失目标,提出一种基于l_1图半监督协同训练的目标跟踪算法。算法首先提取样本的颜色和纹理特征构建两个充分冗余的视图,再以基于l_1图的半监督学习算法取代传统协同训练中的监督学习方法构建单视图中的分类器,提高有限标记样本条件下的分类正确率,然后通过基于负类学习的协同训练算法协同更新两个视图的分类器,最后根据不同视图的相似度分布熵融合各分类器的分类结果实现目标跟踪。实验结果表明,该算法能够有效提高分类器的判别能力,具有良好的跟踪性能。
Ground target tracking by unmanned aerial vehicle （UAV） is more difficult than the traditional video object tracking because the targets have often smaller size and are more blurrecL A tracking algorithm is proposed based on l1 graph-based semi-supervised co-training. Firstly, color and texture features of labeled and unlabeled samples are ex- tracted to construct two sufficient and redundant views. Secondly, l1 graph-based semi-supervised learning is adopted to train two separate classifiers in the co-training framework to replace the traditional supervised learning, which can improve the classifying accuracy under limited labeled training samples. Then, the classifiers from different views teach each other by co-training algorithm based on negative learning. Finally, the confidence distribution entropy of each view from two classifiers is evaluated as its own weight to achieve the tracking. Experimental results show that the proposed algorithm can effectively improve the discriminative ability of the classifier and achieve good tracking performance.
作者
毛盾
邢昌风
满欣
付峰
MAO Dun XING Chang-feng MAN Xin FU Feng(Electronics Engineering College, Naval University of Engineering, Wuhan 430033, China Troops 91208 PLA,Qingdao 266100, China)
出处
《激光与红外》
CAS
CSCD
北大核心
2017年第6期778-782,共5页
Laser & Infrared
基金
国家自然科学基金项目(No.61501484)资助
关键词
目标跟踪
l1图
基于图的半监督学习
多视图协同训练
无人机
object tracking
l1 graph
graph-based semi-supervised learning
multi-view co-training
unmanned aerial vehicle
