Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 159-168.DOI: 10.3778/j.issn.1002-8331.2205-0444

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Siamese Network Long-Term Object Tracking Algorithm Based on Dynamic Template Matching

LI Yang, HOU Ying, LI Jiao, YANG Lin, SHI Huan, HE Shun, ZHANG Shiru   

  1. College of Communication and Information Engineering, Xi’an University of Science and Technology, Xi’an 710054, China
  • Online:2023-09-01 Published:2023-09-01



  1. 西安科技大学 通信与信息工程学院,西安 710054

Abstract: Due to closing to practical application scenarios, long-term object tracking has attracted much attention in recent years. Many long-term tracking algorithms use the object re-detection mechanism to solve the problem of out-of-view and reappearance, but the tracking speed cannot meet the real-time requirements. In the paper a Siamese network long-term object tracking algorithm based on dynamic template matching is proposed. A re-detection mechanism using fast dynamic template matching is designed. When the target loss is detected, template matching method is utilized to roughly predict the object location, and then the accurate target position is further obtained using Siamese network object tracking algorithm. The deformation of the target appearance will seriously affect the template matching positioning accuracy, so the dynamic matching template update strategy is proposed to improve the anti-interference performance. Compared with the state-of-the-art long-term target tracking algorithm on four datasets, the experimental results show that the improved algorithm can not only significantly improve the object tracking performance, but also obtain 40?FPS tracking speed, which can meet the needs of real-time target tracking.

Key words: machine vision, long-term object tracking, deep learning, Siamese network, target re-detection, template matching

摘要: 近年来备受关注的长时目标跟踪更接近实际应用场景,许多算法通过目标重检测机制来解决目标消失与重现问题,但跟踪速度无法满足实时目标跟踪需求,因此提出一种基于模板匹配的孪生网络长时目标跟踪算法。改进算法引入快速动态模板匹配的全局搜索重检测机制,当检测到目标丢失情况,则利用模板匹配进行预测粗定位,再通过孪生网络目标跟踪算法进一步获得目标准确位置。目标外观形变会严重影响模板匹配预测定位的准确度,因此引入动态匹配模板更新策略进一步提高抗干扰性能。在四个数据集上与当前先进的长时目标跟踪算法相比较,实验结果显示改进算法不仅跟踪性能显著提高,并且跟踪速度达到40?FPS左右,能满足实时目标跟踪需求。

关键词: 机器视觉, 长时目标跟踪, 深度学习, 孪生网络, 目标重检测, 模板匹配