谢文杰,
黄鸿,
汤一明
重庆大学光电技术及系统教育部重点实验室 重庆 400044
基金项目:国家自然科学基金(41371338)、重庆市基础与前沿研究计划(cstc2018jcyjAX0093)、重庆市留学人员回国创业创新支持计划(cx2019144)、重庆市研究生科研创新项目(CYB19039, CYB18048)
详细信息
作者简介:刘嘉敏：男，1973年生，副教授，研究方向为图像处理、模式识别
谢文杰：男，1995年生，硕士生，研究方向为图像处理、视频跟踪
黄鸿：男，1980年生，教授，研究方向为流形学习、模式识别和遥感影像智能化处理
汤一明：男，1993年生，博士生，研究方向为模式识别、图像处理、深度学习和视觉跟踪
通讯作者:刘嘉敏 liujm@cqu.edu.cn
中图分类号:TN911.73; TP391.4计量
文章访问数:580
HTML全文浏览量:166
PDF下载量:127
被引次数:0
出版历程
收稿日期:2020-08-05
修回日期:2021-03-20
网络出版日期:2021-04-16
刊出日期:2021-09-16
Spatial and Channel Attention Mechanism Method for Object Tracking
Jiamin LIU,,Wenjie XIE,
Hong HUANG,
Yiming TANG
Key Laboratory of Optoelectronic Technique System of the Ministry of Education, Chongqing University, Chongqing 400044, China
Funds:The National Natural Science Foundation of China (41371338), Chongqing Basic and Frontier Research Program (cstc2018jcyjAX0093), Chongqing Returned Overseas Students’ Entrepreneurship and Innovation Support Program (cx2019144), Chongqing Graduate Research and Innovation Project (CYB19039, CYB18048)
摘要
摘要:目标跟踪是计算机视觉中重要的研究领域之一，大多跟踪算法不能有效学习适合于跟踪场景的特征限制了跟踪算法性能的提升。该文提出了一种基于空间和通道注意力机制的目标跟踪算法(CNNSCAM)。该方法包括离线训练的表观模型和自适应更新的分类器层。在离线训练时，引入空间和通道注意力机制模块对原始特征进行重新标定，分别获得空间和通道权重，通过将权重归一化后加权到对应的原始特征上，以此挑选关键特征。在线跟踪时，首先训练全连接层和分类器层的网络参数，以及边界框回归。其次根据设定的阈值采集样本，每次迭代都选择分类器得分最高的负样本来微调网络层参数。在OTB2015数据集上的实验结果表明：相比其他主流的跟踪算法，该文所提算法获得了更好的跟踪精度，重叠成功率和误差成功率分别为67.6%，91.2%。
关键词:目标跟踪/
深度学习/
空间注意力/
通道注意力/
在线学习
Abstract:Object tracking is one of the important research fields in computer vision. However, most tracking algorithm can not effectively learn the features suitable for tracking scene, which limits the performance improvement of tracking algorithm. To overcome this problem, this paper proposes a target tracking algorithm based on CNN Spatial and Channel Attention Mechanisms (CNNSCAM). The method consists of an off-line training apparent model and an adaptive updating classifier layer. In the offline training, the spatial and channel attention mechanism module is introduced to recalibrate the original features, and the space and channel weights are obtained respectively. The key features are selected by normalizing the weights to the corresponding original features. In online tracking, the network parameters of the full connection layer and classifier layer are trained, and the boundary box regression is used. Secondly, samples are collected according to the set threshold, and the negative sample with the highest classifier score is selected for each iteration to fine tune the network layer parameters. The experimental results on OTB2015 dataset show that compared with other mainstream tracking algorithms, the proposed method achieves better tracking accuracy. The overlap success rate and error success rate are 67.6% and 91.2% respectively.
Key words:Object tracking/
Deep learning/
Spatial attention/
Channel attention/
Online learning
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=39ffd94a-d35c-4f6d-8f50-3efdf3878c9a
