基于二阶池化特征融合的孪生网络目标跟踪算法

Siamese Network Target Tracking Algorithm Based on Second-Order Pooling Feature Fusion

为提升基于孪生网络目标跟踪算法的特征表达能力,获得更好的跟踪性能,提出了一种轻量级的基于二阶池化特征融合的孪生网络目标跟踪算法。首先,使用孪生网络结构获取目标的深度特征;然后,在孪生网络结构的末端并行添加二阶池化网络和轻量级通道注意力,以获取目标的二阶池化特征和通道注意力特征;最后,将目标的深度特征、二阶池化特征和通道注意力特征进行融合,使用融合后的特征进行互相关操作,得到地响应图能很好地区分目标和背景,提高跟踪模型的判别能力,改善目标定位的精度,从而提升跟踪性能。所提算法使用Got-10k数据集进行端到端的训练,并在数据集OTB100和VOT2018上进行验证。实验结果表明,所提算法与基准算法相比,跟踪性能取得了显著提升:在OTB100数据集上,精确度和成功率分别提高了7.5%和5.2%;在VOT2018数据集上,预期平均重叠率(EAO)提高了4.3%。

In order to improve the feature expression ability of the target tracking algorithm based on Siamese network and obtain better tracking performance,a lightweight Siamese network target tracking algorithm based on second-order pooling feature fusion is proposed.First,the Siamese network architecture is used to obtain the deep features of the target;then,the second-order pooling network and the lightweight channel attention are added in parallel at the end of the Siamese network architecture to obtain the second-order pooling features and channel attention features of the target,respectively.Finally,the depth feature of the target,the second-order pooling feature and the channel attention feature are fused,and the fused feature is used for cross-correlation operation,and the obtained response graph can distinguish the target and the background well,and improve the discriminative ability of the model,and improve the accuracy of target positioning,thereby improving target tracking performance.The proposed algorithm only uses the Got-10k dataset for end-to-end training and is validated on the OTB100 and VOT2018 datasets.The experimental results show that the proposed algorithm achieves a significant improvement in tracking performance compared with the benchmark algorithm SiamFC:on the OTB100 dataset,the accuracy and success rate are increased by 7.5%and 5.2%,respectively;on the VOT2018 dataset,the expected average overlap rate increases by 4.3%.