西安电子科技大学学报 ›› 2024, Vol. 51 ›› Issue (3): 46-54.doi: 10.19665/j.issn1001-2400.20231002

• 信息与通信工程 • 上一篇    下一篇

结合模板更新与轨迹预测的孪生网络跟踪算法

贺王鹏(), 胡德顺(), 李诚(), 周悦(), 郭宝龙()   

  1. 西安电子科技大学 空间科学与技术学院,陕西 西安 710071
  • 收稿日期:2023-04-03 出版日期:2024-06-20 发布日期:2023-10-24
  • 作者简介:贺王鹏(1989—),男,副教授,E-mail:hewp@xidian.edu.cn
    胡德顺(1997—),男,西安电子科技大学硕士研究生,E-mail:dshu@stu.xidian.edu.cn
    李 诚(1991—),男,西安电子科技大学博士研究生,E-mail:licheng812@stu.xidian.edu.cn
    周 悦(1998—),女,西安电子科技大学硕士研究生,E-mail:yz@stu.xidian.edu.cn
    郭宝龙(1962—),男,教授,E-mail:blguo@xidian.edu.cn
  • 基金资助:
    国家自然科学基金(52175112);陕西省自然科学基础研究计划资助项目(2023JCYB289);中央高校基本科研业务费专项资金资助项目(ZYTS23102)

Siamese network tracking using template updating and trajectory prediction

HE Wangpeng(), HU Deshun(), LI Cheng(), ZHOU Yue(), GUO Baolong()   

  1. School of Aerospace Science & Technology,Xidian University,Xi’an 710071,China
  • Received:2023-04-03 Online:2024-06-20 Published:2023-10-24

摘要:

目标跟踪一直是计算机视觉领域中重要且富有挑战的问题。为克服目标形变、遮挡或快速移动等因素对跟踪性能的影响,笔者提出一种结合模板更新与轨迹预测的孪生网络跟踪算法。首先,在基于孪生网络跟踪模型中引入模板图像的自适应更新迭代机制,实现对目标表观变化的动态表征,以此提升目标形状或颜色发生变化时的跟踪性能。具体来说,通过对每一帧跟踪结果的分析,判断是否满足更新条件,设计了自适应模板更新的策略,有效地降低了目标模板被污染的可能性。其次,在目标跟踪过程中引入卡尔曼滤波,通过收集跟踪过程中目标位置信息并进行轨迹预测,将前一帧中跟踪算法预测的目标位置信息与轨迹预测的位置信息相融合,得到当前帧搜索区域的裁剪位置,进而实现了离线跟踪与在线学习的结合,进一步解决了目标被遮挡或者快速移动的问题。最后,在VOT2018和LaSOT数据集上验证了该算法在多种复杂场景下的性能表现。实验结果表明,所提算法的跟踪性能超过了大部分其他跟踪算法。

关键词: 深度学习, 目标跟踪, 孪生网络, 模板更新, 轨迹预测, 卡尔曼滤波

Abstract:

Object tracking is an active and challenging issue in the field of computer vision.To tackle the problem that a target may suffer from deformation,occlusion and fast motion during the tracking process,a novel Siamese network tracking algorithm is proposed,with emphasis on template updating and trajectory prediction.First,an effective template updating mechanism is introduced to the Siamese network tracking model that adaptively represents the variation of target appearance.This mechanism could further improve the tracking performance when the target suffers from shape or color deformation.Specifically,by analyzing the tracking results of each frame to determine whether the update conditions are met,an adaptive template update strategy is designed,effectively reducing the possibility of template contamination.Second,the Kalman filter is utilized to collect the target position information and predict the motion trajectory.By fusing the object position information predicted by the tracking algorithm in the previous frame with the position information predicted by the trajectory,the cropping position of the search area in the current frame is obtained,which further solves the problem of the object being occluded or moving quickly by combining offline tracking and online learning.Extensive experiments on the VOT2018 and LaSOT datasets verify that the tracking performance of the proposed approach exceeds that obtained by other state-of-the-art algorithms under various complex scenarios.

Key words: deep learning, object tracking, Siamese network, template updating, trajectory prediction, Kalman filtering

中图分类号: 

  • TP391