基于特征融合与双模板嵌套更新的孪生网络跟踪算法

为提高全卷积孪生网络SiamFC在复杂场景下的识别和定位能力，提出一种基于多响应图融合与双模板嵌套更新的实时目标跟踪算法。使用深度ResNet-22替换AlexNet作为骨干网络以提升网络特征提取性能，建立强识别能力的骨干语义分支。在ResNet-22的浅层使用高分辨率特征，构造强定位能力的浅层位置分支，计算并融合两个分支响应。通过高置信度的双模板嵌套更新机制对两个分支的模板进行更新，以适应目标的外观和位置变化。在OTB2015和VOT2016数据集上的实验结果表明，与基于SiamFC、SiamDW等的目标跟踪算法相比，该算法在目标快速移动、遮挡等复杂场景下跟踪效果更稳定，并且运行速度达到34 frame/s，满足实时性要求。

Tracking Algorithm Using Siamese Network Based on Feature Fusion and Dual-Template Nested Update

In order to improve the recognition and positioning performance of the fully convolutional Siamese network(SiamFC) in complex scenarios,a real-time visual tracking algorithm with fused multiple response graphs and dual-template nested update mechanism is proposed.The algorithm employs the deep network,ResNet-22,to replace AlexNet as the backbone network for stronger feature extraction ability,and the semantic branch of backbone with enhanced recognition ability is built.The high-resolution feature is used in the shallow layer of ResNet-22 to construct the shallow position branch with strong positioning ability.Then the responses of the two branches are calculated and fused.In addition,the templates of the two branches are updated by using a high-confidence dual-template nested update mechanism to adapt to the changes in the appearance and position of the target.Experimental results on the datasets of OTB2015 and VOT2016 show that the algorithm is more stable than tracking algorithms based on SiamFC,SiamDW and other networks in the scenarios with Fast Motion(FM) and Occlusion(OCC).At the same time,the algorithm runs at the speed of 34 frame/s,providing required real-time performance.