| 摘要: |
| 在当前的目标跟踪领域，现有的基于分割的算法没有充分利用目标的长距离依赖信息和各个特征层的不同特性，前背景判别能力不强，对目标的多尺度估计不足。针对此问题，提出了自适应特征融合模块和混合域注意力模块，以提高网络对目标的多尺度估计能力和对目标的前背景辨别能力，并将其集成到当前基于视频分割的算法中，提出了一种新的目标跟踪算法，在各大公开数据集上的实验结果证明其达到了领先水平。 |
| 关键词: 目标跟踪 深度学习 特征融合 注意力机制 分割算法 |
| DOI： |
|
| 基金项目:国家科技重大专项（2017ZX03001004）；重庆市研究生教育教学改革研究项目（yjg212024） |
|
| A target segmentation and tracking algorithm based on feature fusion and attention mechanism |
| WANG Shiyan,ZHANG Qingsong,LEI Guofang,ZHANG Jiangshan |
| (School of Communication and Information Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China) |
| Abstract: |
| In current target tracking field,the existing algorithms do not make full use of the long-distance dependence information of the target and the different characteristics of each feature layer,and the front background discrimination ability is not strong,the multi-scale estimation of target is underestimated,too.For above problems,an adaptive feature fusion module and a hybrid domain attention module are proposed to improve the network’s multi-scale estimation ability of the target and the ability to discriminate the front background of the target.By integrating the modules into the current video segmentation-based algorithm,a new target tracking algorithm is proposed.Test results in all major public data sets prove that the proposed algorithm reaches a leading level. |
| Key words: target tracking deep learning feature fusion attention mechanism segmentation algorithm |
