基于堆栈式消噪自编码机的分块目标跟踪（英文）
发布时间：2018-08-26 06:52
【摘要】：在视觉目标跟踪系统中,特征的表达和提取是重要的组成部分.本文提出基于多个自编码机网络相联合的特征提取机,通过对输入数据进行一定程度的重组,采用深度学习的理论对其局部特征进行描述并对结果进行联合决策.结合该网络结构,本文提出一种融合局部特征的深度信息进行目标跟踪的算法.将输入图像分块使得大量的乘法运算转化为加法和乘法的混合运算,相对于全局的特征表达,大幅降低了运算复杂度.在跟踪过程中,目标候选区的各分块权重能够根据相应网络的置信度进行自适应的调整,提升了跟踪器对光照变化、目标姿态和遮挡的适应.实验表明,该跟踪算法在鲁棒性和跟踪速度上表现优秀.
[Abstract]:The representation and extraction of features is an important part of visual target tracking system. In this paper, a feature extraction machine based on multiple self-coding machine networks is proposed. By reorganizing the input data to a certain extent, the local features are described by using the theory of depth learning and the results are jointly determined. Combined with the network structure, this paper proposes an algorithm for target tracking based on the depth information of local features. When the input image is partitioned into blocks, a large number of multiplication operations are transformed into mixed operations of addition and multiplication. Compared with the global feature representation, the computational complexity is greatly reduced. In the tracking process, each block weight of the target candidate area can be adjusted adaptively according to the confidence degree of the corresponding network, and the adaption of the tracker to the illumination change, target attitude and occlusion is improved. The experimental results show that the proposed tracking algorithm performs well in robustness and tracking speed.
【作者单位】： 空军工程大学信息与导航学院;解放军第93716部队;
【基金】：Supported by National Natural Science Foundation of China(61473309) Natural Science Foundation of Shaanxi Province(2015JM6269,2016JM6050)
【分类号】：TP391.41
本文编号：2204069
[Abstract]:The representation and extraction of features is an important part of visual target tracking system. In this paper, a feature extraction machine based on multiple self-coding machine networks is proposed. By reorganizing the input data to a certain extent, the local features are described by using the theory of depth learning and the results are jointly determined. Combined with the network structure, this paper proposes an algorithm for target tracking based on the depth information of local features. When the input image is partitioned into blocks, a large number of multiplication operations are transformed into mixed operations of addition and multiplication. Compared with the global feature representation, the computational complexity is greatly reduced. In the tracking process, each block weight of the target candidate area can be adjusted adaptively according to the confidence degree of the corresponding network, and the adaption of the tracker to the illumination change, target attitude and occlusion is improved. The experimental results show that the proposed tracking algorithm performs well in robustness and tracking speed.
【作者单位】： 空军工程大学信息与导航学院;解放军第93716部队;
【基金】：Supported by National Natural Science Foundation of China(61473309) Natural Science Foundation of Shaanxi Province(2015JM6269,2016JM6050)
【分类号】：TP391.41
【相似文献】
相关期刊论文 前5条
1 赵蕴龙;王博识;张凯;董钊;张磊;;充分考虑节点编码机会的编码感知路由协议[J];应用科学学报;2014年01期
2 王晓鸣;TST700袖珍式话音编码机[J];通信保密;1991年02期
3 马义德,张宗念,余英林;H.263+部分可选模式的研究[J];兰州大学学报;2000年04期
4 姜波，，胡坚;一个有趣而全新的无线世界即将来临[J];国外科技动态;1994年12期
5 ;[J];;年期
相关会议论文 前2条
1 罗庆生;韩宝玲;吴瑞松;;液晶显示技术在新型多功能喷墨编码机中的应用[A];'99全国包装与食品加工技术研讨会论文集[C];1999年
2 张小卫;;中国喷墨编码机市场概况[A];'99全国包装与食品加工技术研讨会论文集[C];1999年
相关硕士学位论文 前2条
1 肖巍;无线网络中基于即时译码网络编码的研究[D];南京邮电大学;2016年
2 代思;无线网络中编码机会发现方法的研究[D];宁波大学;2014年
本文编号：2204069
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2204069.html
