计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (23): 237-245.DOI: 10.3778/j.issn.1002-8331.2208-0208

• 图形图像处理 • 上一篇    下一篇

融合低通滤波器的孪生网络跟踪算法

杨晓强,刘文昊   

  1. 西安科技大学 计算机科学与技术学院,西安 710000
  • 出版日期:2023-12-01 发布日期:2023-12-01

Siamese Network Tracking Algorithm of Fused Low Pass Filter

YANG Xiaoqiang, LIU Wenhao   

  1. College of Computer Science and Technology, Xi’an University of Science and Technology, Xi’an 710000, China
  • Online:2023-12-01 Published:2023-12-01

摘要: 针对深层次网络中的填充操作会破坏网络严格的平移不变性,提出了一种改进的孪生神经网络目标跟踪算法(BsSiamCAR)。在SiamCAR基础上,用ResNet-50网络对模板图像和搜索区域进行特征提取时,采用低通滤波器和最大池化相融合的策略,改善平移不变性对深层次网络的影响,提高深层次网络的稳定性。为了解决目标跟踪中背景干扰、尺度变化等复杂场景中的问题,引入通道注意力,注意力机制能够有选择性地突出对跟踪有利的特征通道,增强算法在复杂环境中的鲁棒性。实验结果表明,BsSiamCAR在OTB100、VOT2018、UAV123数据集上较SiamRPN和SiamCAR等多种算法在成功率和精度上均有提升,跟踪速度达到43 FPS。

关键词: 目标跟踪, 孪生网络, 低通滤波器, 通道注意力, 最大池化

Abstract: An improved Siamese network target tracking algorithm(BsSiamCAR) is proposed for deep-level networks in which the padding operation can destroy the strict translational invariance of the network. Based on SiamCAR, when using ResNet-50 for target extraction of template images and search images, a low-pass filter and max-pooling are fused to improve the effect of translation invariance on the deep network and improve the deep network. For the problems in complex scenes such as background interference and scale change in target tracking, channel attention is introduced, which can selectively highlight the feature channels favorable to tracking and enhance the robustness of the algorithm in complex environments. The experimental results show that BsSiamCAR has improved in success rate and accuracy over various algorithms such as SiamRPN and SiamCAR on OTB100, VOT2018, and UAV123 datasets, and the tracking speed reaches 43 FPS.

Key words: target tracking, Siamese network, low-pass filters, channel attention, max-pooling