Abstract: For the siamese network tracking framework, the object representation capability of features extracted from the backbone network is essential for the object tracking performance. To improve the ability of features to characterize the object, a siamese tracking algorithm based on feature-aware enhancement is proposed. Based on TrDiMP algorithm, it uses a feature-aware enhancement module to enhance the features extracted by the ResNet-50 backbone network. It first utilizes self-attention to independently enhance the features from the template branch and the search branch and enhances the features from both branches based on cross-attention, establishing the correlation between template features and search features to suppress the interference of object-independent features and further enhance the representation capability of features for the object. Extensive experiments show that, compared with several state-of-the-art siamese trackers, the proposed algorithm achieves superior accuracy and success rates on 5 benchmark datasets, namely, OTB100, UAV123, LaSOT, GOT-10k, and VOT2018, especially in complex scenarios such as similarity interference, scale changes, occlusion, and other complex scenarios.

Key words: object tracking, siamese network, feature-aware enhancement, attention

摘要： 对于孪生网络跟踪框架，骨干网络提取特征的目标表征能力对目标跟踪性能至关重要。为了提升特征对目标的表征能力，提出一种基于特征感知增强的孪生跟踪算法。其以TrDiMP算法为基础，对ResNet-50骨干网络提取的特征采用特征感知增强模块进行增强，该模块基于自注意力对模板分支和搜索分支的特征进行自我增强，并基于交叉注意力对两个分支进行关联增强，建立模板特征和搜索特征之间的目标依赖关系，抑制与目标无关的特征干扰，进而提升特征对目标的表征能力。基于OTB100、UAV123、LaSOT、GOT-10k和VOT2018五个基准数据集的大量实验表明，与几种主流的孪生跟踪器相比，该算法取得了更优的精确度和成功率，尤其在相似性干扰、尺度变化、遮挡等复杂场景中鲁棒性更优。

关键词: 目标跟踪, 孪生网络, 特征感知增强, 注意力