结合随机掩膜与特征融合的孪生网络目标跟踪

Combining random soft mask and feature fusion for visual object tracking

针对全卷积孪生神经网络在尺度变化、变形、背景相似等情况容易出现跟踪失败的问题,提出了一种结合随机掩膜与特征融合的孪生网络目标跟踪算法.主干网络使用特征提取能力更强的VGGNet网络代替AlexNet网络,添加随机软掩膜来模拟复杂环境;在模板图像添加3分支注意力机制模块,将主干网络的第4-1层和5-1层进行特征融合;使用大规模数据对网络进行端到端训练.在5个公开测试集的实验表明,该算法在尺度变化、变形、背景相似等复杂环境下仍具有良好的跟踪性能,并且在NVIDIA RTX2070S上跟踪速度达到54FPS,满足实时性要求.

In order to solve the problem that the fully-convolutional siamese network is easy to fail to track in complex scenarios,such as scale variation,deformation,background clutters and so on,a siamese network algorithm combining random soft mask and feature fusion is proposed.Firstly,the backbone network uses the VGGNet network with stronger feature extraction ability to replace AlexNet network.Secondly,a random soft mask is added to simulate a complex environment.Then,an improved triplet attention module is added to the template image.Finally,it uses the features of layer 4-1 and layer 5-1 of the backbone network.Experiments on five publicly available benchmark data sets show that the algorithm still has good tracking performance in complex environment,such as scale variation,deformation,background clutters,etc.,and the tracking speed reaches 54 FPS on NVIDIA RTX2070S,which meets the real-time require monts.