类残差更新策略的核相关滤波目标跟踪

Kernel Correlation Filtering Target Tracking Based on Quasi-Residual Updating Strategy

针对相关滤波类目标跟踪算法存在的两个问题,提出了结合特征融合与类残差更新策略的核相关滤波实时目标跟踪算法。首先,为了解决相关滤波目标跟踪算法中使用相同系数结合不同种类特征的问题,根据平均相关峰值能量提出一种自适应特征融合的背景感知相关滤波器;其次,将特征融合的背景感知相关滤波器与贝叶斯分类器通过集成学习整合成一个强跟踪器;最后,针对相关滤波器使用高风险的更新策略的情况,提出类残差的更新策略与基于阶跃函数的学习率更新跟踪模型以降低跟踪模型漂移的风险。将该算法在OTB2013与TC128测试集上与其他9个先进的算法进行了比较,实验结果表明,该算法在OTB2013数据集中的精确度与成功率分别为0.875和0.652,排名第一,具有很好的鲁棒性。

In order to deal with two existing problems of correlation filter based target tracking algorithms a real-time target tracking algorithm based on kernel correlation filtering is proposed which combines feature fusion with quasi-residual updating strategy.Firstly to solve the problem that correlation filters employ the same coefficient to fuse different kinds of features a correlation filter based on context-awareness with adaptive feature fusion is proposed according to the average correlation peak energy.Furthermore the correlation filter based on context-awareness and adaptive feature fusion is combined with a Bayes classifier to construct a robust tracker by ensemble learning.Finally focusing on the use of high-risk updating strategy in correlation filters an updating strategy that is similar to deep residual networks and an adaptive learning rate updating model based on step function are proposed to prevent tracking model from drifting.The tracker proposed in this paper is compared with another 9 state-of-the-art trackers on OTB2013 and TC128 benchmarks.The experimental result on OTB2013 benchmark is that the proposed tracker ranks first on precision(0.875)and success rate(0.652)which indicates that the proposed tracker is robust and effective.