基于参数自适应与模板更新的孪生网络跟踪算法

Siamese network tracking algorithm based on parameter adaptive and template updating

孪生网络跟踪算法在跟踪过程中网络参数固定,跟踪模板仅仅使用第1帧给定的目标,这导致算法的鲁棒性较差。为此,提出基于参数自适应(PA)与模板更新的孪生网络跟踪算法。首先,利用通道注意力和空间注意力对目标特征进行调整,提高网络对跟踪目标的关注度;其次,利用滤波器参数更新策略滤除背景的干扰,提高网络对当前目标的辨识能力;最后,增加与主网络平行的子网络,通过更新子网络的跟踪模板,使网络能适应目标的变化。在VOT 2018、VOT 2019 2个标准数据集上进行测试,期望重叠率(EAO)分别达到0.455和0.331,验证了本算法的有效性。

The network parameters of the Siamese network tracking algorithm are fixed during the tracking process,and the tracking template is only from the first frame,which make the robustness of algorithm poor.Therefore,a Siamese network tracking algorithm based on parameter adaptive(PA)and template updating is proposed.Firstly,the target feature is adjusted by channel attention and spatial attention to improve the attention of network to tracking target;secondly,the filter parameter update strategy is used to filter out the interference of background,which leads to identify the current target;finally,a subnetwork parallel to the main network is added,and the network can adapt to the change of the target by updating the tracking template.The expected average overlap(EAO)reaches 0.455 and 0.331 respectively on the VOT2018 and VOT2019 benchmarks,which verifies the effectiveness of the algorithm.