基于自适应深度稀疏网络的在线跟踪算法
发布时间：2018-10-11 10:25
【摘要】：视觉跟踪中,高效鲁棒的特征表达是解决复杂环境下跟踪漂移问题的关键。该文针对深层网络预训练复杂费时及单网络跟踪易漂移的问题,在粒子滤波框架下,提出一种基于自适应深度稀疏网络的在线跟踪算法。该算法利用Re LU激活函数,针对不同类型目标构建了一种具有自适应选择性的深度稀疏网络结构,仅通过有限标签样本的在线训练,就可得到鲁棒的跟踪网络。实验数据表明:与当前主流的跟踪算法相比,该算法的平均跟踪成功率和精度均为最好,且与同样基于深度学习的DLT算法相比分别提高了20.64%和17.72%。在光照变化、相似背景等复杂环境下,该算法表现出了良好的鲁棒性,能够有效地解决跟踪漂移问题。
[Abstract]:In visual tracking, efficient and robust feature representation is the key to solve the problem of tracking drift in complex environment. Aiming at the complex and time-consuming pre-training of deep network and the easy drift of single network tracking, this paper proposes an online tracking algorithm based on adaptive deep sparse network under the framework of particle filter. Using the Re LU activation function, the algorithm constructs a kind of self-adaptive and selective deep sparse network structure for different types of targets. The robust tracking network can be obtained only by the online training of finite tag samples. The experimental data show that the average tracking success rate and accuracy of the algorithm are the best compared with the current mainstream tracking algorithms, and the DLT algorithm based on the same depth learning is increased by 20.64% and 17.72% respectively. In complex environments such as illumination variation and similar background, the proposed algorithm is robust and can effectively solve the drift tracking problem.
【作者单位】： 空军工程大学信息与导航学院;
【基金】：国家自然科学基金(61473309) 陕西省自然科学基础研究计划项目(2015JM6269,2016JM6050)~~
【分类号】：TP391.41
本文编号：2263851
[Abstract]:In visual tracking, efficient and robust feature representation is the key to solve the problem of tracking drift in complex environment. Aiming at the complex and time-consuming pre-training of deep network and the easy drift of single network tracking, this paper proposes an online tracking algorithm based on adaptive deep sparse network under the framework of particle filter. Using the Re LU activation function, the algorithm constructs a kind of self-adaptive and selective deep sparse network structure for different types of targets. The robust tracking network can be obtained only by the online training of finite tag samples. The experimental data show that the average tracking success rate and accuracy of the algorithm are the best compared with the current mainstream tracking algorithms, and the DLT algorithm based on the same depth learning is increased by 20.64% and 17.72% respectively. In complex environments such as illumination variation and similar background, the proposed algorithm is robust and can effectively solve the drift tracking problem.
【作者单位】： 空军工程大学信息与导航学院;
【基金】：国家自然科学基金(61473309) 陕西省自然科学基础研究计划项目(2015JM6269,2016JM6050)~~
【分类号】：TP391.41
【相似文献】
相关期刊论文 前1条
1 程健,曾以亮;IPv6在产品特性在线跟踪中的应用[J];机械与电子;2004年07期
，本文编号：2263851
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2263851.html
