Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (16): 177-186.DOI: 10.3778/j.issn.1002-8331.2204-0454

• Graphics and Image Processing • Previous Articles     Next Articles

Tracking Algorithm Combining Multi-Level Feature Fusion and Efficient Attention

YAO Zhuangze, ZENG Bi, LIN Zhentao, JIANG Chunling, DENG Bin   

  1. School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2023-08-15 Published:2023-08-15

结合多级特征融合和高效注意力的跟踪算法

姚壮泽,曾碧,林镇涛,江春灵,邓斌   

  1. 广东工业大学 计算机学院,广州 510006

Abstract: In complex scenes such as illumination change, size change, and motion blur, the existing target tracking algorithms are easy to fail. To solve these problems, some tracking algorithms try to introduce global attention to enhance feature expression for improving accuracy, but it needs to consume a lot of computing resources. This paper proposes an object tracking algorithm SiamEff(Siamese efficient network) that combines multi-level feature fusion and efficient attention, which enables global attention to improve accuracy while only using a small number of computing resources. Firstly, an anchor-free mechanism is employed to avoid the sensitivity of anchor-based tracking algorithms to a large number of hyperparameters. Then, an attention mechanism is introduced, which is more efficient than conventional global attention and greatly reduces memory and computational requirements, low-level and high-level multi-level features are fused, the feature information at different levels is used to improve tracking performance. Finally, centrality constraints are introduced to reduce the interference of low-quality prediction boxes. Tested on the public test platforms OTB100 and VOT2018, the results show that SiamEff tracking performance is better than mainstream tracking algorithms, more robust and accurate in various complex scenarios, and the tracking speed reaches 150 FPS.

Key words: object tracking, anchor-free, attention mechanism, center-ness, multi-level feature

摘要: 在光照变化、形状变化、运动模糊等复杂场景下,现有的目标跟踪算法很容易跟踪失败,为了解决这些问题,一些跟踪算法尝试引入全局注意力来增强特征表达以提高准确率,但这需要消耗大量的计算资源。提出一种结合多级特征融合和高效注意力的目标跟踪算法SiamEff(Siamese efficient network),使全局注意力能在耗费少量计算资源的前提下提高准确率。使用无锚框机制以避免基于锚框的跟踪算法对大量超参的敏感问题;引入一种比传统全局注意力更加高效的注意力机制,大大降低内存和运算量的需求,并融合了低层和高层的多级特征,充分利用不同层次的特征信息提高跟踪性能;引入了中心度约束,以减少低质量预测框的干扰。在公开测试平台OTB100和VOT2018上进行了测试,结果表明SiamEff的跟踪性能优于主流跟踪算法,在各种复杂场景下更加鲁棒和准确,且跟踪速度达到150?FPS。

关键词: 目标跟踪, 无锚框, 注意力机制, 中心度约束, 多级特征