计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (12): 112-121.DOI: 10.3778/j.issn.1002-8331.2110-0200

• 模式识别与人工智能 • 上一篇    下一篇

引入轻量注意力的孪生神经网络目标跟踪算法

洪培钦,罗灵鲲,刘冰,方元,胡士强   

  1. 1.上海交通大学 航空航天学院,上海 200240
    2.中国航空无线电电子研究所,上海 200241
  • 出版日期:2022-06-15 发布日期:2022-06-15

Siamese Neural Network Target Tracking Algorithm with Lightweight Attention

HONG Peiqin, LUO Lingkun, LIU Bing, FANG Yuan, HU Shiqiang   

  1. 1.School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China
    2.China National Aeronautical Radio Electronics Research Institute, Shanghai 200241, China
  • Online:2022-06-15 Published:2022-06-15

摘要: 针对目标跟踪算法在各种场景下很难做到准确率和实时性平衡的问题,提出了一种引入轻量注意力的孪生神经网络(siamese neural network)目标跟踪算法,称为SiamNL。SiamNL算法使用基于深度级卷积(depth-wise convolution)的交叉相关运算,降低了网络的参数量和运算量,提升了算法的实时性。同时,SiamNL使用Non-Local注意力网络编码模板图特征和搜索图特征,对特征进行了自注意力和互注意力的运算,有效提升了算法的准确率。在VOT2016、VOT2018、OTB100等公开数据集上的测试结果表明,SiamNL算法优于主流的目标跟踪算法,更有效地平衡了准确率和实时性。

关键词: 目标跟踪, 孪生神经网络, 轻量注意力, 实时性

Abstract: Aiming at the problem that it is difficult to balance the accuracy and real-time performance of target tracking tasks across various scenarios, a novel siamese neural network target tracking algorithm, namely, SiamNL is proposed. The proposed SiamNL algorithm leverages cross-correlation operations based on depth-wise convolution, and thereby reducing the amount of network parameters and operations, and improving the real-time performance of the algorithm within the unified deep learning framework. Additionally, SiamNL borrows the merits of the non-local attention module to encode template graph features and search graph features, therefore self-attention and cross-attention operations are performed on the features, so as to effectively improve the accuracy of the algorithm. The experimental results demonstrate that the proposed SiamNL lifts the performance of the state-of-the-art approaches with a large margin crossing three popular datasets, e.g., VOT2016, VOT2018, and OTB100, while the solid experimental discussion verifies the contribution of the proposed SiamNL in balancing the accuracy and real-time performance.

Key words: target tracking, siamese neural network, lightweight attention, real-time