计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (22): 190-198.DOI: 10.3778/j.issn.1002-8331.2006-0402

• 模式识别与人工智能 • 上一篇    下一篇

基于LCT+的自适应抗遮挡目标跟踪算法

陈富健,谢维信,夏婷   

  1. 深圳大学 ATR国防科技重点实验室,广东 深圳 518060
  • 出版日期:2021-11-15 发布日期:2021-11-16

Adaptive Anti-occlusion Target Tracking Algorithm Based on LCT+

CHEN Fujian, XIE Weixin, XIA Ting   

  1. Key Laboratory of ATR National Defense Science and Technology, Shenzhen University, Shenzhen, Guangdong 518060, China
  • Online:2021-11-15 Published:2021-11-16

摘要:

在目标跟踪过程中,目标遮挡往往会造成跟踪器的性能下降,从而导致目标丢失。针对这一问题,提出一种基于LCT+核相关滤波的自适应抗遮挡目标跟踪算法。该算法在LCT+核相关滤波算法的基础上进行改进,利用双跟踪器自适应对目标进行跟踪,即根据两个跟踪器的输出响应值大小选择最优跟踪器跟踪目标;利用支持向量机自适应重新检测目标,即根据目标丢失帧的数量自适应调整检测框范围的大小;最后采用颜色直方图匹配的方法进一步验证预测的目标。相比原算法,所提算法采取双跟踪器自适应跟踪机制和支持向量机自适应重检测机制,有效避免了目标跟丢。在OTB50和OTB100两个大型基准数据集上对算法进行验证,结果表明该算法在距离精度和成功率的评估指标上都优于一些主流算法,并且在抗遮挡方面具有较高的精度和较强的鲁棒性。

关键词: 目标跟踪, 核相关滤波, 自适应抗遮挡, 双跟踪器, 支持向量机(SVM)

Abstract:

In the process of target tracking, the occlusion of the target reduces the performance of the tracker, resulting in the loss of the target. To solve this problem, this paper proposes an adaptive anti-occlusion target tracking algorithm based on LCT+ kernel correlation filter. The algorithm is improved on the basis of LCT+ kernel correlation filtering algorithm, which proposes a strategy of using two trackers to adaptively track the target. According to the output response value of the two trackers, the optimal tracker is selected to track the target. In addition, a strategy for re-detecting targets by using support vector machines is proposed. The detection range is adaptively adjusted according to the number of frames lost of the target. Finally, the predicted target is verified by using color histogram matching. Compared with the original algorithm, the algorithm in this paper adopts a dual tracker mechanism to adaptively track the target and a support vector machine mechanism to adaptively re-detect the target, which effectively avoids the loss of the target. The proposed algorithm is validated in two large benchmark data sets of OTB50 and OTB100. The results indicate that the proposed algorithm is superior to some mainstream algorithms in the evaluation indexes of distance precision and success rate. In terms of target anti-occlusion, it has higher accuracy and strong robustness.

Key words: target tracking, kernel correlation filtering, adaptive anti-occlusion, dual tracker, Support Vector Machine(SVM)