DCNN深度特征与交替方向乘子的相关滤波跟踪

Correlation Filter Tracking Integrated DCNN Depth Feature with Alternating Direction of Multipliers

针对采用相关滤波的判别式目标跟踪遇到的瓶颈问题:由于目标快速移动引起边界效应,使得相关滤波器在学习与更新过程中可能会引入错误,最终错误的累积将导致跟踪失败。在采集深度学习特征与样本相似性度量的基础上,提出一种引入交替方向乘子方法的改进相关滤波目标跟踪算法,选择DCNN深度特征有效地表征待跟踪目标的初始状态,通过在线分类过程中样本相似性比对与半监督学习,辅助解决相关滤波器在学习过程中存在的自学习问题。所提目标跟踪算法特别适合训练样本为持续获得的、同时存储空间较小的机器学习过程,提高目标在快速运动与部分遮挡等复杂情况下的跟踪成功率,针对VOT2016标准测试视频的实验表明:当目标面临快速运动时,对比CN、SAMF、STC算法,所提DA-CFT跟踪算法将跟踪成功率分别由60.4%～73.4%、67.2%～82.9%、80.9%～88.1%提升至85.6%～91.0%。

This paper is aiming at the bottleneck problem of discriminative target tracking using correlation filter. On the period of learning and updating on correlation filter, error is likely to be induced into filter, and fatal cumulation will finally cause tracker inefficiency. It is attributed to the boundary effect caused by the rapid movement of the target. Based on acquisition of depth learning features and sample similarity measure, an improved correlation filter target tracking algorithm with alternating direction method of multipliers is proposed in this paper. DCNN depth features are selected to effectively represent the initial state of the object to be tracked. The algorithm uses sample similarity matching with semi supervised learning in online classification process. The above methods assist in solving the self-learning problem of the correlation filter in the learning process. The proposed object tracking algorithm is especially suited to machine learning process where samples are continually acquired and memory storage is limited. On the complicated scenes such as object speed drastically changing and object partially blocked, etc., success tracking rate is updated by importing the proposed object tracking algorithm. When target vehicle is partially blocked, the experiments on standard testing videos involving VOT2016 demonstrate that success tracking rate of the proposed DA-CFT tracking algorithm is raised to 85.6%~91.0%,compared with 60.4%~73.4% of CN algorithm, 67.2%~82.9% of SAMF algorithm and 80.9%~88.1% of STC algorithm.