CSAM:用于单目标跟踪的卷积自注意力模块  

CSAM:Convolutional Self-attention Module for Single Object Tracking

作　　者：郭崇 张杨洋 张文波[1] 朱宏博 尹震宇 GUO Chong;ZHANG Yangyang;ZHANG Wenbo;ZHU Hongbo;YIN Zhenyu(School of Information Science and Engineering,Shenyang Ligong University,Shenyang 110159,China;School of Computer Science and Engineering,Northeastern University,Shenyang 110169,China;Shenyang Institute of Computing Technology,Chinese Academy of Sciences,Shenyang 110168,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159 [2]东北大学计算机科学与工程学院,沈阳110169 [3]中国科学院沈阳计算技术研究所,沈阳110168

出　　处：《小型微型计算机系统》2024年第11期2702-2709,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金青年项目(62102272)资助.

摘　　要：为了在目标发生遮挡、形变、尺度变化和背景干扰等场景下准确地跟踪目标,注意力机制被广泛应用于特征抽取模块,以选择性地关注重要特征和抑制无关特征.然而,现有的注意力机制只考虑了通道特征层与空间特征点之间的局部或全局关系,没有对特征进行融合建模.本文针对复杂跟踪场景提出了一种基于卷积神经网络和自注意力机制的卷积自注意力模块(Convolutional Self-Attention Module,CSAM),该模块能够以注意力加权方式解决前景遮蔽、非刚性形变、快速尺度变化与相似特征背景干扰问题.经过实验验证,引入卷积自注意力模块的孪生网络能够显著地提升跟踪器的性能,在跟踪问题基准(Benchmark)数据集OTB100上以平均重叠率、跟踪成功率与准确率作为评判指标,相比基准模型分别提升了9.2%、2.2%与2.9%.通过进一步的消融实验证明了本文提出并引入的适用于孪生网络跟踪框架的卷积自注意力模块能够有效地提升特征辨识度,对比先进方案兼顾单目标跟踪性能和实时性,能够在大多数实时复杂跟踪场景实现轻量化部署.To achieve accurate object tracking in complex scenarios with foreground occlusion,deformation,varying scale,and background interference,attention mechanisms are commonly utilized in feature extraction modules to selectively focus on significant features while suppressing irrelevant ones.However,existing attention mechanisms only consider the local or global relationships between channel feature maps and spatial feature points,without modeling the fusion of features.In this paper,we propose a Convolutional Self-Attention Module(CSAM)based on convolutional neural networks and self-attention mechanisms to deal with complex tracking scenarios.This module can solve foreground occlusion,non-rigid deformations,rapid scale changes and similar background interference problems in an attention-weighted manner.Experimental results show that the Siamese Network with the introduced CSAM significantly improves the performance of the tracker.On the benchmark dataset OTB100,evaluated using average overlap,success rate,and precision as performance metrics,the proposed model outperforms the baseline model by 9.2%,2.2%,and 2.9%respectively.Through further ablation experiments,it is proved that the convolutional self-attention module proposed and introduced in this paper is suitable for the siamese network tracking framework,which can effectively improve the feature recognition.Compared with the advanced scheme,it takes into account the single-object tracking performance and real-time performance,and can be used in most real-time complex tracking scenarios realize lightweight deployment.

关 键 词：单目标跟踪 复杂跟踪场景 孪生网络 自注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]