基于孪生网络的轻量化单目标跟踪算法研究  

Research on Lightweight Single Object Tracking Algorithm Based on Siamese Network

作　　者：黎清华 张雪秋 张雨珊 LI Qinghua;ZHANG Xueqiu;ZHANG Yushan(College of Mechanical Engineering,Chongqing University of Technology,Chongqing,400054,China)

机构地区：[1]重庆理工大学机械工程学院,重庆400054

出　　处：《现代信息科技》2025年第3期84-89,共6页Modern Information Technology

摘　　要：针对目前多数基于孪生网络的单目标跟踪算法模型参数大,难以在嵌入式平台等移动端设备部署的问题,设计了一种基于SiamRPN改进的轻量化孪生网络单目标跟踪算法。算法使用裁剪过的MobileNetV3替换AlexNet作为特征提取网络,降低算法的参数量;融合了MobileNetV3的高低层特征,整合不同尺度的特征信息,增强算法对目标多尺度变化的适应能力;使用GhostConv替换原网络模型头部的普通卷积,使用深度可分离卷积进行相关运算,在保证特征提取能力的同时进一步减少了参数量。在OTB数据集上进行测试的结果表明:改进后的算法在准确率上提高了1.6%,成功率提高了3.8%,模型参数量为2.9 MB,减少了98.3%,证明改进后的目标跟踪算法模型具有显著的性能优势。Aiming at the problem that most of the Single Object Tracking algorithms based on Siamese Network have large model parameters and are difficult to deploy on mobile devices such as embedded platforms,a Single Object Tracking algorithm of improved lightweight Siamese Network based on SiamRPN is designed.The algorithm uses the clipped MobileNetV3 to replace AlexNet as the feature extraction network to reduce the number of parameters of the algorithm.It integrates the highlevel and low-level features of MobileNetV3,integrates the feature information of different scales,and enhances the adaptability of the algorithm to the multi-scale changes of the object.GhostConv is used to replace the ordinary convolution of the head of the original network model,and the Depthwise Separable Convolution is used with for correlation operation,which further reduces the number of parameters while ensuring the feature extraction ability.The test results on the OTB dataset show that the improved algorithm improves the accuracy by 1.6%,and the success rate by 3.8%.The number of model parameters is 2.9 MB,which is reduced by 98.3%.It is proved that the improved object tracking algorithm model has significant performance advantages.

关 键 词：SiamRPN 单目标跟踪 轻量化网络 特征融合 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] TP183[自动化与计算机技术—计算机科学与技术]