融合注意力机制的孪生网络目标跟踪算法研究

Siamese Network Tracking Algorithms for Hierarchical Fusion of Attention Mechanism

在全卷积孪生网络跟踪算法(SiamFC)的基础上,提出一种融合注意力机制的孪生网络目标跟踪算法。在网络模板分支,通过融合注意力机制,由神经网络学习模板图像的通道相关性和空间相关性,进而增大前景贡献,抑制背景特征,提升网络对正样本特征的辨别力;同时,使用VggNet-19网络提取模板图像的浅层特征和深层特征,两种特征自适应融合。在OTB2015和VOT2018数据集上得到的实验结果表明,与SiamFC相比,所提算法能够更好地应对运动模糊、目标漂移和背景多变等问题,取得了更高的准确率和成功率。

Based on the full-convolution Siamese network tracking algorithm SiamFC,this paper proposes a Siamese network target tracking algorithm fused attention mechanism.In the template branch,the neural network can learn the channel correlation and the spatial correlation of the template image through the attention mechanism fusion,thus increasing the foreground contribution,suppressing the background features,and improving the discrimination of network to positive samples features.Meanwhile,the VggNet-19 network is used to extract the shallow and deep features of the template image,the two features fuse adaptively.The experimental results on the datasets of OTB2015 and VOT2018 demonstrate that compared with SiamFC,the proposed algorithm can more effectively deal with the tracking problems,such as motion blur,target drift and background clutter,obtains higher accuracy and success rate.