当前位置:主页 > 科技论文 > 软件论文 >

基于卷积特征的核相关自适应目标跟踪

发布时间:2018-05-29 18:06

  本文选题:目标跟踪 + 卷积特征 ; 参考:《中国图象图形学报》2017年09期


【摘要】:目的针对现实场景中跟踪目标的快速运动、旋转、尺度变化、遮挡等问题,提出了基于卷积特征的核相关自适应目标跟踪的方法。方法利用卷积神经网络提取高、低层卷积特征并结合本文提出的核相关滤波算法计算并获得高底两层卷积特征响应图。采用Coarse-to-Fine方法对目标位置进行估计,在学习得到1维尺度核相关滤波器估计尺度的基础上实时更新高低两层核相关滤波器参数,以实现自适应的目标跟踪。结果实验选取公开数据集中的典型视频序列进行跟踪,测试了算法在目标尺度发生变化、遮挡、旋转等复杂场景下的跟踪性能并与多种优秀的跟踪算法在平均中心误差、平均重叠率等指标上进行了定量比较,在Singer1、Car4、Jogging、Girl、Football以及MotorRolling视频图像序列上的中心误差分别为8.71、6.83、3.96、3.91、4.83、9.23,跟踪重叠率分别为0.969、1.00、0.967、0.994、0.967、0.512。实验结果表明,本文算法与原始核相关滤波算法相比,平均中心位置误差降低20%,平均重叠率提高12%。结论采用卷积神经网络提取高低两层卷积特征,高层卷积特征用于判别目标和背景,低层卷积特征用于预测目标位置并通过Coarse-to-Fine方法对目标位置进行精确的定位,较好地解决了由于目标的旋转和尺度变化带来的跟踪误差大的问题,提高了跟踪性能并能够实时更新学习。在目标尺度发生变化、遮挡、光照条件改变、目标快速运动等复杂场景下仍表现出较强的鲁棒性和适应性。
[Abstract]:Aim to solve the problems of fast moving, rotation, scale change, occlusion and so on, a kernel correlation adaptive target tracking method based on convolution feature is proposed. Methods the high and low level convolution features were extracted by convolution neural network and the response map of the high bottom two-layer convolution feature was obtained by combining the kernel correlation filtering algorithm proposed in this paper. The Coarse-to-Fine method is used to estimate the location of the target. Based on the learning of the estimation scale of the 1-dimensional kernel correlation filter, the parameters of the high and low two layers correlation filter are updated in real time to achieve adaptive target tracking. Results the typical video sequences in the open data set were selected for tracking, and the tracking performance of the algorithm in complex scenes such as target scale change, occlusion, rotation and so on was tested, and the average center error of the algorithm was compared with that of many excellent tracking algorithms. Quantitative comparison was made on the average overlap rate. The central errors in Singer 1 / Car4 Jogging GirlsFootball and MotorRolling video sequences were 8.71 / 6.83 / 3.963.91 / 4.839.23, respectively, and the tracking overlap rates were 0.9691.000.9670.9940.967/ 0.967/ 0.512, respectively. The experimental results show that compared with the original kernel correlation filtering algorithm, the average center position error is reduced by 20% and the average overlap rate is increased by 12%. Conclusion the convolution neural network is used to extract the convolution feature of high and low layers, the high level convolution feature is used to distinguish the target and the background, and the lower level convolution feature is used to predict the target position and to locate the target position accurately by Coarse-to-Fine method. The problem of large tracking error caused by target rotation and scale change is well solved, and the tracking performance is improved and the learning can be updated in real time. It still shows strong robustness and adaptability in complex scenes such as the change of target scale, occlusion, light condition change, fast moving of target and so on.
【作者单位】: 中北大学计算机与控制工程学院;
【基金】:山西省自然科学基金项目(2013011017-6)~~
【分类号】:TP391.41

【相似文献】

相关期刊论文 前10条

1 H.Meerkamm,St.Sander,R.-D.Kasan,陈巍;一个建立特征图书库的方案[J];工程设计学报;2001年04期

2 陈佳;;基于眼睛特征图的眼睛检测系统的设计与实现[J];南阳师范学院学报;2008年06期

3 魏永伟;曹娟;汪国昭;;平面三次混合双曲多项式曲线的特征图判别[J];计算机辅助设计与图形学学报;2010年05期

4 孙正兴,,丁秋林;基于特征的CAD/CAM集成技术评述与研究[J];计算机应用;1994年06期

5 祝国旺;钟毅芳;周济;余俊;;零件形状特征图树分解原理及应用[J];交通与计算机;1993年02期

6 段冠新;;在SolidWorks中库特征的创建和使用[J];机械工人.冷加工;2006年02期

7 葛江华,隋秀凛,邹世浩;基于特征的电站阀门CAD/CAPP/CAM集成系统[J];哈尔滨理工大学学报;2002年02期

8 李净净;张基宏;梁永生;;视觉注意模型中权值估计与特征整合[J];深圳信息职业技术学院学报;2012年03期

9 徐志;关宏志;严海;陈二慧;;基于图像显著性特征的交通标志注视点预测方法[J];武汉理工大学学报(交通科学与工程版);2014年01期

10 许本胜;黄美发;苏庆勇;王灿;;基于特征几何变动的三维公差综合建模方法研究[J];制造业自动化;2014年06期

相关硕士学位论文 前3条

1 杨凡;基于层次特征的视觉注意模型研究[D];华中科技大学;2016年

2 魏艾;基于全局和局部深度特征的图像重排序方法研究[D];中国科学技术大学;2017年

3 陈超;基于个人微博特征的Timeline生成研究[D];内蒙古科技大学;2015年



本文编号:1951885

资料下载
论文发表

本文链接:https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/1951885.html


Copyright(c)文论论文网All Rights Reserved | 网站地图 |

版权申明:资料由用户434ac***提供,本站仅收录摘要或目录,作者需要删除请E-mail邮箱bigeng88@qq.com