计算机科学与探索 ›› 2019, Vol. 13 ›› Issue (1): 83-96.DOI: 10.3778/j.issn.1673-9418.1710052

• 人工智能与模式识别 • 上一篇    下一篇

自适应模板更新的鲁棒目标跟踪算法

刘  芳,黄光伟+,路丽霞,王洪娟,王  鑫   

  1. 北京工业大学 信息学部,北京 100124
  • 出版日期:2019-01-01 发布日期:2019-01-09

Robust Target Tracking Algorithm for Adaptive Template Updating

LIU Fang, HUANG Guangwei+, LU Lixia, WANG Hongjuan, WANG Xin   

  1. Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
  • Online:2019-01-01 Published:2019-01-09

摘要: 针对受复杂背景、光照以及目标尺度变化等因素的影响,目标模板更新精度不高,导致跟踪算法鲁棒性差的问题,提出了一种基于深度特征和模板更新的自适应粒子滤波目标跟踪方法。首先对跟踪目标进行仿射变换;然后构造一个12层的卷积神经网络来提取跟踪目标及其仿射变换的深度特征得到目标模板和候选模板,并以此构建候选模板库;其次采用粒子滤波算法跟踪目标,将预测结果与候选模板库中的模板进行匹配,确定新的目标模板并自适应更新候选模板库。实验结果表明,该算法在遮挡、光照、尺度变化、目标旋转和复杂背景的恶劣条件下仍能稳定地跟踪目标,与其他7种先进算法在18组测试视频中进行比较,具有更高的目标跟踪精度和更强的鲁棒性。

关键词: 目标跟踪, 卷积神经网络, 粒子滤波, 模板更新

Abstract: An algorithm of adaptive particle filter target tracking based on depth feature and template updating is proposed in view of the problem that the accuracy of the target template update is not high, resulting in poor tracking algorithm robust problem, due to the influence of factors such as complex background, illumination and target scale change. Firstly, the tracking target is done affine transformation, and then the target template and the candidate template are obtained by the 12-layers convolution neural network constructed to extract the depth feature of the tracking target and its affine transformation, and the candidate template library is constructed on this account. Secondly, using the particle filter tracking algorithm, the prediction target is matched with the template in the candidate template library, determining the new target template and updating adaptively the candidate template library. The experimental results show that the algorithm can track the target stably under the bad conditions of occlusion, illumination, scale change, target-rotated and complex background. Compared with the other seven adv-anced algorithms in eighteen sets of test videos, the algorithm has higher target tracking precision and stronger robustness.

Key words: target tracking, convolution neural network, particle filter, template update