基于ShuffleNetV2的孪生网络目标跟踪算法  

Siamese Network Target Tracking Algorithm Based on ShuffleNetV2

作　　者：杨国宇 董琴 YANG Guoyu;DONG Qin(School of Mechanical Engineering,Yancheng Institute of Technology,Yancheng 224001,China)

机构地区：[1]盐城工学院信息工程学院,江苏盐城224001

出　　处：《软件导刊》2024年第12期255-261,共7页Software Guide

摘　　要：孪生网络跟踪算法可将跟踪问题转换为相似性匹配问题,然而多数算法无法在移动端或算力不足的嵌入式设备中实现工程应用。为此,提出一种基于孪生网络的轻量级跟踪算法,选择可以在移动端使用的ShuffleNetV2作为核心网络。针对原网络的不足,提出消除padding层影响、修改激活函数、采用上采样、修改步长4种优化操作,同时引入注意力机制进一步加强特征之间的联系。在OTB100和UAV123数据集上进行仿真实验,结果表明,与现有跟踪算法相比,所提算法综合性能表现优异,同时在面对形变、低分辨率、尺度变换等多种复杂因素带来的影响时有较好的鲁棒性。Siamese network tracking algorithms can transform tracking problems into similarity matching problems,but most algorithms cannot be implemented in engineering applications on mobile devices or embedded devices with insufficient computing power.To this end,a lightweight tracking algorithm based on siamese networks is proposed,selecting ShuffleNetV2 as the core network that can be used on mobile devices.Aiming at the shortcomings of the original network,four optimization operations are proposed:eliminating the influence of padding layer,modifying activation function,adopting upsampling,and modifying step size.At the same time,attention mechanism is introduced to further strengthen the connection between features.Simulation experiments were conducted on the OTB100 and UAV123 datasets,and the results showed that compared with existing tracking algorithms,the proposed algorithm has excellent comprehensive performance.At the same time,it has good robustness in the face of various complex factors such as deformation,low resolution,and scale transformation.

关 键 词：目标跟踪 ShuffleNetV2 孪生网络 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]