摘要
Struck在面对目标尺度变化的场景时,跟踪结果的精度会降低,甚至出现丢失目标的情况。为解决这一问题,提出了一种在分类判别器中引入尺度变量的解决方案,在提高算法对尺度变化的鲁棒性的同时提高算法的实时性。对算法的改进主要是在分类判别器添加尺度变量,让分类器学习目标的尺度信息。同时在采样过程中,利用支持向量机中的判别函数实现目标位置预测,并减少计算量、提高实时性。改进后的算法在Car4、Girl、Walking、Basketball三个视频序列测试集上进行测试,运行平台为I7处理器、Win10系统,实时性和精准度都有明显的提升。其中,在Girl测试序列上距离精度提升最为明显,在阈值为26个像素时,达到31.2%。理论分析和实验证明,改进的算法有效地解决了传统Struck算法对目标尺度变化鲁棒性不好的问题,实时性也有所提高。
When Struck faces a scene with a change in target scale,the accuracy of the tracking results will be reduced,and even the loss of the target will occur.In order to solve this problem,a solution to introduce scale variables into the classifier is proposed to improve the real-time performance of the algorithm while improving the robustness of the algorithm to scale changes.The improvement of the algorithm is mainly to add scale variables in the classifier to let the classifier learn the scale information of the target.At the same time,in the sampling process,the discriminant function in the support vector machine is used to achieve the target position prediction,and the calculation amount is reduced and the real-time performance is improved.The improved algorithm is tested on the three mainstream video sequence test sets of Car4,Girl,Walking and Basketball.The running platform is I7 processor and Win10 system,and the real-time and accuracy are obviously improved.Among them,the distance accuracy improvement is most obvious in the Girl test sequence,reaching 31.2%when the threshold is 26 pixels.Theoretical analysis and experiments show that the improved algorithm effectively solves the problem that the traditional Struck algorithm is not robust to target scale changes,and the real-time performance is also improved.
作者
蔡雨
黄学功
张志安
宋峻
苏冶
CAI Yu;HUANG Xue-gong;ZHANG Zhi-an;SONG Jun;SU Ye(School of Mechanical Engineering,Nanjing University of Science&Technology,Nanjing 210000,China;726th Research Institute of CSIC,Shanghai 201108,China;Heilongjiang Xinnuo Robot Automation Co.,Ltd.,Harbin 150000,China)
出处
《电子设计工程》
2020年第16期1-6,共6页
Electronic Design Engineering
基金
国家自然科学基金(11472008,11772160,11202206)。