孪生压缩激励全卷积网络的目标跟踪方法

吉林大学学报(理学版) ›› 2021, Vol. 59 ›› Issue (2): 325-332.

孪生压缩激励全卷积网络的目标跟踪方法

刘庆强¹, 张福禄¹, 张瑶瑶², 张晨雨³

1. 东北石油大学电气信息工程学院, 黑龙江大庆 163318； 2. 东北石油大学计算机与信息技术学院, 黑龙江大庆 163318； 3. 安徽农业大学经济管理学院, 合肥 230061

收稿日期:2020-04-21 出版日期:2021-03-26 发布日期:2021-03-26
通讯作者: 张福禄 E-mail:heshang_181@163.com

Target Tracking Method Based on Siamese Squeeze and Excitation Fully Convolutional Networks

LIU Qingqiang¹, ZHANG Fulu¹, ZHANG Yaoyao², ZHANG Chenyu³

1. School of Electrical Engineering and Information, Northeast Petroleum University, Daqing 163318, Heilongjiang Province, China;
2. School of Computer and Information Technology, Northeast Petroleum University, Daqing 163318, Heilongjiang Province, China;
3. College of Economics and Management, Anhui Agricultural University, Hefei 230061, China

Received:2020-04-21 Online:2021-03-26 Published:2021-03-26

摘要/Abstract

摘要： 为提升目标跟踪的准确性并保证其实时性, 提出一种基于改进孪生全卷积网络的新方法——孪生压缩激励全卷积网络(siamese squeeze and excitation fully convolutional networks, Siam-SEFC). Siam-SEFC通过添加具有少量参数的压缩激励网络结构融合空间通道信息, 为跟踪对象添加空间信息, 并通过调整训练数据尺度进行尺度不定的数据增强, 提取多尺度特征, 有效提升目标跟踪的准确性. 为提升多尺度训练速度, 网络采用单一尺度预训练的权重进行初始化. 与MDNet,SENet,DAT三种算法相比, Siam-SEFC在保证目标跟踪准确性的同时具有实时性；而与Siamese-FC相比, Siam-SEFC跟踪准确性提升了2.2%, 参数量仅增加1.01%, 且未损失实时性, 验证了改进方案的有效性.

关键词: 目标跟踪, 孪生全卷积网络, 实时性, 多尺度特征, 数据增强

Abstract: In order to improve the accuracy of the target tracking and achieve real-time tracking, we proposed a new method based on improved siamese fully convolutional networks: siamese squeeze and excitation fully convolutional networks (Siam-SEFC). Siam-SEFC fused spatial channel information by adding squeeze and excitation networks structure with a small number of parameters, added spatial information for the tracking object, and adjusted the training data scale for variable-scale data enhancement to extract multi-scale features, which effectively improved the accuracy of target tracking. Compared with the three algorithms of MDNet, SENet and DAT, Siam-SEFC has real-time performance while ensuring the accuracy of target tracking. Compared with Siamese-FC, the tracking accuracy of Siam-SEFC is improved by 2.2%, the number of parameters is only increased by 1.01%, and there is no loss of real-time performance, which verifies the effectiveness of the improved scheme.

Key words: target tracking, siamese fully convolutional networks, real-time, multi-scale feature, data enhancement

中图分类号:

TP391

刘庆强, 张福禄, 张瑶瑶, 张晨雨. 孪生压缩激励全卷积网络的目标跟踪方法[J]. 吉林大学学报(理学版), 2021, 59(2): 325-332.

LIU Qingqiang, ZHANG Fulu, ZHANG Yaoyao, ZHANG Chenyu. Target Tracking Method Based on Siamese Squeeze and Excitation Fully Convolutional Networks[J]. Journal of Jilin University Science Edition, 2021, 59(2): 325-332.

[1]	车颖, 冯皛, 郑宏亮. 基于卷积神经网络的超声造影图像去噪方法[J]. 吉林大学学报(理学版), 2021, 59(5): 1256-1259.
[2]	朱新丽, 才华, 寇婷婷, 杜冬晖, 孙俊喜. 行人多目标跟踪算法[J]. 吉林大学学报(理学版), 2021, 59(5): 1161-1170.
[3]	张彩丽, 刘广文, 詹旭, 史浩东, 才华, 李英超. 基于Mask R-CNN的多目标跟踪算法[J]. 吉林大学学报(理学版), 2021, 59(3): 609-618.
[4]	张尧, 才华, 曹露, 王冰雪, 陈广秋. 基于字典学习改进的时空上下文算法[J]. 吉林大学学报(理学版), 2019, 57(06): 1442-1448.
[5]	董立岩, 王雪松, 王朝阳, 李永丽. 基于区域活跃用户的好友推荐和位置推荐算法[J]. 吉林大学学报(理学版), 2018, 56(6): 1441-1446.
[6]	张睿萍, 马宗梅. 基于Hadoop平台的大数据图像分类机制[J]. 吉林大学学报(理学版), 2018, 56(5): 1206-1212.
[7]	李秀娟. 一种抗遮挡的运动目标跟踪算法[J]. 吉林大学学报(理学版), 2017, 55(05): 1207-1212.
[8]	王江涛. 基于背景差分和粒子滤波的运动目标跟踪算法[J]. 吉林大学学报(理学版), 2015, 53(05): 999-1005.
[9]	王欣, 赵连义, 薛龙. 基于主成分分析的粒子滤波器目标跟踪方法[J]. J4, 2012, 50(06): 1156-1162.
[10]	张博, 申铉京. 一种基于目标定位的背景建模与视觉跟踪方法[J]. J4, 2010, 07(4): 617-623.
[11]	臧雪柏，赵亚莉, 王秋实，郑凯，陈鹏, 赵宏伟. Linux 2.6内核的中断机制[J]. J4, 2008, 46(03): 494-498.
[12]	张忠波, 马驷良, 董险峰. 基于局部特征分析与最优化匹配的人脸识别算法[J]. J4, 2005, 43(01): 59-63.