Computer Engineering and Applications ›› 2020, Vol. 56 ›› Issue (19): 132-138.DOI: 10.3778/j.issn.1002-8331.1907-0266

Previous Articles     Next Articles

Error Refinement Method of Target Tracking Based on Multi-scale Region Proposal

ZHANG Hongwei, LI Xiaoxiang, ZHU Bin, MA Qi   

  1. School of Electronic Countermeasure, National University of Defense Technology, Hefei 230037, China
  • Online:2020-10-01 Published:2020-09-29



  1. 国防科技大学 电子对抗学院,合肥 230037


With the introduction of deep learning technology into the field of visual target tracking, the accuracy and robustness of the target tracking have greatly improved. However, in the actual scene of target tracking with low altitude UAV, the complex situations, such as camera jitter, a large number of occlusions, changes of shooting perspective and focal length, etc. , make the tracking algorithm face great challenges. Most of recent algorithms are based on the assumption that the appearance of the target changes slowly; they do not have the ability to detect and repair the drift(tracking error)in the tracking process. In order to solve this problem, an error refinement method of target tracking based on multi-scale region proposal is proposed. In the offline stage, a large number of labeled target samples are used to train the tracking refinement model based on multi-scale region proposal method, which can obtain prior knowledge of different kinds of targets. In the online stage, kernel correlation filter-based tracking algorithm is used to track the target in real time. According to the self-adaptive evaluation of the correlation response confidence, the tracking refinement model is used to initialize the position of target irregularly, so as to avoid the loss of target due to the error accumulation. Experimental results on UAV aerial datasets show that the proposed algorithm can correct the tracking drift problem accurately when targets undergo heavy deformation. Compared with the state-of-art algorithms, the success rate and the precision of target tracking are improved by 14.3% and 3.1%, respectively.

Key words: visual tracking, convolutional neural network, kernel correlation filter, tracking drift, cumulative error



关键词: 视觉跟踪, 卷积神经网络, 核相关滤波, 跟踪漂移, 误差累积