基于无锚的轻量化孪生网络目标跟踪算法

Lightweight Siamese Network Target Tracking AlgorithmBased on Ananchor Free

针对运算资源受限条件下难以实现高精度、高帧率跟踪的问题,提出一种基于无锚的轻量化孪生网络目标跟踪算法.首先使用修改的轻量级网络MobileNetV 3作为主干网络提取特征,在保持深度特征表达能力的同时减小网络的参数量和计算量;然后对传统互相关操作,提出图级联优化的深度互相关模块,通过丰富特征响应图突出目标特征重要信息;最后在无锚分类回归预测网络中,采用特征共享方式减少参数量和计算量以提升跟踪速度.在两个主流数据集OTB2015和VOT2018上进行对比实验,实验结果表明,该算法相比于SiamFC跟踪器有较大的精度优势,并且在复杂跟踪场景下更具鲁棒性,同时跟踪帧率可达175帧/s.

Aiming at the problem that it was difficult to achieve high-precision and high frame rate tracking under limited computing resources,we proposed a lightweight siamese network target tracking algorithm based on ananchor free.Firstly,the modified lightweight network MobileNetV3 was used as the backbone network to extract features,and reduced parameters and computation of the network while maintaining deep feature expression capability.Secondly,for traditional cross-correlation operation,we proposed deep cross-correlation module for graph cascading optimization,which highlighted important information of target features through rich feature response graphs.Finally,feature sharing was used to reduce paramet ers and computation to improve tracking speed in the anchor classification regression prediction network.Comparative experiments were conducted on two mainstream datasets OTB2015 and VOT2018,the experimental results show that the algorithm has a significant accuracy advantages compared to SiamFC tracker,and is more robust in complex tracking scenes.At the same time,the tracking frame rate can reach 175 frames/s.