结合注意力与特征融合的目标跟踪

全卷积孪生网络通过相似性学习解决目标跟踪问题，其算法受到了越来越多的关注.为了提取更有判别力的目标特征，提升跟踪的精确度和鲁棒性，提出了一种结合注意力机制与特征融合的目标跟踪模型.首先，将第一帧和当前帧的前一帧结合作为目标模板，利用共享的特征提取网络提取目标模板和当前帧的多个卷积层的特征；然后，对于目标模板的多层卷积特征，结合通道注意力机制处理，提升模板特征的判别力；最后，目标模板的特征与当前帧的特征进行互相关计算，得到响应图，从而获取预测目标在当前帧中的位置和尺度.最终实验结果表明，与几个先进的跟踪模型相比，提出的目标跟踪模型获得了比较有竞争力的性能.

Object Tracking Combined with Attention and Feature Fusion

The full-convolutional Siamese network solves the problem of object tracking through similarity learning, and the algorithm has received more and more attention. In order to extract more discriminative object features and improve the accuracy and robustness of tracking, an object tracking model combining attention mechanism and feature fusion is proposed. Firstly, the first frame and the previous frame of the current frame are combined as target templates, and the features from multiple convolution layers of the target templates and the current frame are extracted by using the shared feature extraction network. Furthermore, for the multi-layer convolution features of the target templates, the channel attention mechanism is adopted to improve the discriminative power of the template features. Finally, the features of the target templates are cross-correlated with the features of the current frame to obtain response map, thereby obtaining the position and scale of the predicted object in the current frame. The final experimental results show that compared with several advanced tracking models, the proposed object tracking model achieves relatively competitive performance.