基于局部敏感直方图的压缩跟踪.pdfVIP

214 2016 ，52（2 ） Computer Engineering and Applications 计算机工程与应用 基于局部敏感直方图的压缩跟踪 钱 凯，陈秀宏，孙百伟 QIAN Kai, CHEN Xiuhong, SUN Baiwei 江南大学 数字媒体学院，江苏 无锡 214122 School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China QIAN Kai, CHEN Xiuhong, SUN Baiwei. Compressive tracking based on locality sensitive histograms. Computer Engineering and Applications, 2016, 52 （2 ）：214-218. Abstract ：As traditional tracking algorithm based on compressive sensing is failed to track target stably when illumina- tion and target position get seriously changed, a reformative target tracking algorithm based on locality sensitive histo- grams is proposed. By calculating the locality sensitive histograms, extracting illumination invariant features, the jointed features are used in compressive tracking. Results of tests on variant video sequences show that the proposed algorithm has advantages over compressive tracking and multiple instance learning tracking in target positions and lightings changing heavily. Obtained results satisfy the requirements of real-time tracking. Key words ：compressive tracking; drift; locality sensitive histograms; illumination invariant features; multiple instance learning 摘 要：压缩跟踪在光照发生剧烈变化和目标姿势变化较大时容易出现漂移甚至跟丢现象。针对此缺陷，提出基于 局部敏感直方图的压缩跟踪。通过计算局部敏感直方图，提取光照不变特征，联合压缩跟踪中使用的特征得到更优 的特征。对不同视频序列的跟踪结果表明，与压缩跟踪和多示例学习跟踪算法相比，提出的算法在目标姿势发生较 大变化和光照变化剧烈的情况下能够实现稳定的跟踪，并且满足实时性要求。 关键词：压缩跟踪；漂移；局部敏感直方图；光照不变特征；多示例学习 文献标志码：A 中图分类号：TP391 doi ：10.3778/j.issn.1002-8331.1401-0294 1 引言 更新带有噪声或错误样本时容易引起跟踪漂移。于是， 视频跟踪由于其在运动分析、行为识别、监控和人 Grabner 等[9 又提出了半在线boosting 方法解决漂移问 机交互等方面的广泛应用已成为最活跃的研究课题之 题。Babenko 等[ 0 提出了多实例学习（Multiple Instance 一[ 。基于外观模型跟踪算法可以分为生成模型跟踪[2 6 Learning ，MIL ）跟踪方法，即把有歧义的所有正样本和 [7 0 负样本放入包中学习鉴别模型。但是，以上算法的自适 和鉴别模型跟踪 。生成模型用目标模板或子空间模 板在图像中寻找最大可能的区域完成跟踪。