结合可变形卷积与全局信息的目标跟踪算法

Target tracking algorithm combining deformable convolution and global information

本文提出了一种以区域候选孪生网络(SiamRPN)为基础并结合可变形卷积与全局信息的目标跟踪算法。首先,使用计算花销适中的主干网络提升模型的特征提取能力;其次,采用全局上下文注意力模块提升全局信息建模能力,在相似度计量部分设计可变形互相关模块聚合模板特征与搜索特征;最后,采用多层特征融合策略充分挖掘深层语义信息与浅层定位信息,使目标的定位和分类更加准确。实验结果表明:该算法优于参与对比的主流跟踪器,在OTB100和VOT2016两个目标跟踪数据集中成功率和EAO指标分别提升了5.3%和8.5%,且跟踪速度达到68 fps,达到超实时跟踪,证明所提出算法的有效性。

A target tracking algorithm based on regional candidate twin networks(SiamRPN),and combining deformable convolution and global information.Firstly,backbone network with moderate computation cost is used to improve the feature extraction ability of the model.Secondly,the global context attention module is used to improve the ability of global information modeling.In the part of similarity measurement,the deformable cross-correlation module is designed to aggregate template features and search features.Finally,the multi-layer feature fusion strategy is adopted to thoroughly mine the deep semantic information and shallow positioning information,so that to make target localization and classification more accurate.Experimental results show that this algorithm is obviously better than mainstream trackers in the comparision.In two datasets of target tracking OTB 100 and VOT 2016,the success rate and EAO are improved by 5.3 % and 8.5 % respectively,and the tracking speed reaches 68 fps,realize ultra real-time tracking,which proves the validity of the proposed algorithm.