基于tr-PCBAMSiam的小目标跟踪算法
Small target tracking algorithm based on tr-PCBAMSiam
苏冲 1雷斌 1蒋林 1汪杰 1李港1
作者信息
- 1. 武汉科技大学 机械自动化学院,湖北 武汉 430081;武汉科技大学 机器人与智能系统研究院,湖北 武汉 430081
- 折叠
摘要
无人机在目标跟踪过程中,存在分辨率低、运动模糊、目标遮挡、目标密集、相似目标干扰等问题,导致算法跟踪精度下降.针对这一问题,在SiamRPN的基础上提出tr-PCBAMSiam,即一种基于混合注意力的聚合残差连接、transformer互相关运算以及基于无锚框的区域回归网络的目标跟踪算法.将该算法与其他目标跟踪算法在OTB100数据集上进行对比,在精度和成功率方面,与SiamRPN算法相比分别有 6.9%和 8%的提升;在LaTOT数据集上与SiamRPN相比,精度和成功率分别有 13.1%和 8.5%的提升.
Abstract
During the target tracking process of drones,there are problems such as low resolution,motion blur,target occlusion,dense target,interference of similar targets,etc.,resulting in a decrease in algorithm tracking accuracy.To solve this problem,tr-PCBAMSiam was proposed on the basis of SiamRPN,which was a target tracking algorithm of aggregated residual connection based on mixed attention,transformer cross-correlation operation and region regression network without anchor frame.The algorithm of this study was compared with other target tracking algorithms on the OTB100 dataset.In terms of accuracy and success rate,there were 6.9%and 8%improvements respectively compared with the SiamRPN algorithm;compared with SiamRPN on the LaTOT dataset,the accuracy and the success rates were increased by 13.1%and 8.5%respectively.
关键词
目标跟踪/聚合残差连接/transformer互相关/孪生网络/特征融合Key words
target tracking/aggregate residual connection/transformer cross-correlation/twin network/feature fusion引用本文复制引用
基金项目
国家重点研发计划(2019YFB1310000)
湖北省自然科学基金(2018CFB626)
武汉市应用基础前沿项目(2019010701011404)
机器人与智能系统研究院开放基金(F201803)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据