计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (20): 181-187.DOI: 10.3778/j.issn.1002-8331.2105-0125

• 模式识别与人工智能 • 上一篇    下一篇

基于DiMP模型迁移重组的RGBD目标跟踪

邱守猛,谷宇章,袁泽强   

  1. 1.中国科学院 上海微系统与信息技术研究所 仿生视觉系统实验室,上海 200050
    2.中国科学院大学,北京 100049
  • 出版日期:2022-10-15 发布日期:2022-10-15

DiMP Based Transfer Learning and Model Recomposion for RGBD Tracking

QIU Shoumeng, GU Yuzhang, YUAN Zeqiang   

  1. 1.Bionic Vision System Laboratory, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China
    2.University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2022-10-15 Published:2022-10-15

摘要: 深度图可以提供运动目标所处的三维空间结构信息,因此可以用来提升跟踪性能。但目前缺少基于RGBD的目标跟踪数据集,无法直接训练RGBD输入下的深度学习跟踪器。对此,提出了一种基于知识对齐的模型迁移重组算法,可以方便地将在其他RGBD任务上训练得到的模型迁移到基于DiMP的跟踪算法上来,并且对于不同的跟踪对象不需要重新计算迁移参数。另外,针对深度图信息不稳定的问题,提出了一种高效的平滑稳定算法。在VOTRGBD数据集上的实验结果表明,迁移融合后的特征可以显著提升目标和背景之间的判别性,有效提升跟踪器的性能。

关键词: 模型迁移, 深度信息, 信息融合, 目标跟踪

Abstract: The depth map provides information about the three-dimensional spatial structure of the moving objects, so it can be used to improve tracking performance. However, the current lack of RGBD-based target tracking datasets makes it impossible to directly train deep learning trackers based on RGBD input. In this regard, a model transfer learning algorithm based on knowledge alignment is proposed. Models trained on other RGBD tasks can be easily transferred to the DiMP-based object tracking algorithm, and there is no need to recalculate the transfer parameters for different tracking objects. In addition, a smoothing algorithm is proposed to solve the problem of instability of depth information. The experimental results on the VOTRGBD dataset show that the transferred and fused features can significantly improve the discrimination between the target and the background, and effectively improve the performance of the tracker.

Key words: transfer learning, depth information, information fusion, object tracking