基于特征融合的双分支模板更新跟踪算法

Dual-branch template update tracking algorithm based on feature fusion

为了提高孪生网络SiamFC++的跟踪精确度和尺度适应性,提出了一种具有特征融合和双分支模板动态更新机制的实时目标跟踪算法。针对跟踪精确度,在骨干网络的浅层设计特征融合分支从而提高特征提取能力。另外,使用平均峰值相关能量(APCE)方法判断分类模板是否更新从而提高分类能力,改善了遮挡和形变时的跟踪效果。针对尺度适应性,使用交并比梯度比值法和响应图方差率判断回归模板是否更新,增强了快速移动和尺度变化的适应性。为了保证实时性,将两个分支的更新过程嵌套起来形成双分支模板动态更新机制。在数据集OTB2015和VOT2018上的结果表明,该算法较其他几种算法跟踪效果更稳定,能更好应对快速移动和遮挡等场景,同时该算法达到了62 fps,满足实时性要求。

In order to improve the tracking accuracy and scale adaptability of SiamFC++,this paper proposes a real-time target tracking algorithm with feature fusion and dual-branch template dynamic update mechanism.For tracking accuracy,feature fusion branches are designed in the shallow layer of the backbone network to improve feature extraction ability.In addition,the APCE method is used to determine whether the classification template is updated to improve the classification ability and improve the tracking effect during occlusion and deformation.For the scale adaptability,the IOU gradient ratio method and the response graph variance rate are used to determine whether the regression template is updated,which enhances the adaptability of rapid movement and scale changes.In order to ensure real-time performance,the update processes of the two branches are nested to form a dual-branch template dynamic update mechanism.The results on the data sets OTB2015 and VOT2018 show that the algorithm has a more stable tracking effect than other algorithms,and can better deal with scenes such as fast movement and occlusion.At the same time,the algorithm reaches 62 frames per second,which meets real-time requirements.