基于规范化注意力机制的孪生单目标视觉追踪

Siamese single target visual tracking based on normalization-attention mechanism

为了解决基于全卷积孪生网络(SimaFC)在目标发生遮挡、光照变化、快速运动时发生的追踪失败问题,提出了基于规范化注意力机制的孪生网络(NAMSiamNet),通过引入特征增强模块和规范化注意力机制(NAM)来提升追踪网络的性能.首先利用特征增强模块将特征信息进行二次萃取,保留重要参数的同时去掉大量无效参数,得到映射优化后的特征子空间;随后加载最新提出的高效的、轻量级的基于规范化的注意力机制,用于抑制不太显著的特征信息从而突出显著特征,有助于追踪器克服背景信息干扰.在特征增强模块和NAM注意力机制的双重作用下,追踪器能高效地通过模板图像与搜索区域的相似性度量准确无误地捕捉目标对象,避免发生模型漂移.在目前广泛采用的目标追踪公开数据集OTB2015上,将研究提出的算法与基线算法进行了对比实验,数值结果表明,所提出的追踪网络模型在精确率和成功率上分别提升了4%和2.3%,并且在消融实验中充分验证了研究提出的网络模型结构的合理性.

In order to solve the tracking failure problem based on the fully convolutional Siamese network(SimaFC)when the target is occluded,illumination changes,and fast motion,a Siamese network based on the Normalization-based Attention Mechanism(NAMSiamNet)was proposed,which improve the performance of the tracking network through Normalization-based Attention Mechanism(NAM)with Feature Enhancement Module.Firstly,the Feature Enhancement Module is used to extract the feature information twice,retain important parameters,remove a large number of invalid parameters,and obtain the feature subspace after mapping optimization;Then,the newly proposed efficient and lightweight NAM is loaded to suppress the less salient feature information,so as to highlight the salient features,which helps the tracker not to be disturbed by the background information.Under the dual action of the Feature Enhancement Module and NAM attention mechanism,the tracker can efficiently capture the target object through the similarity measure between the template image and the search area,and avoid model drift.On the widely used target tracking public data set OTB2015,the algorithm proposed in this paper is compared with the baseline algorithm.The experiment shows that the proposed tracking network model plays a positive role,and improves the accuracy and success rate by 4%and 2.3%respectively,and the ablation experiments fully verify the rationality of the network structure model proposed in this paper.