摘要
针对上下文感知相关滤波目标跟踪算法中,上下文背景样本等值权重训练,对背景信息滤波过于平滑的问题,提出了一种自适应上下文感知相关滤波算法,同时为了解决目标遮挡的问题,引入一种新的遮挡判定指标。首先,提取目标上下左右4个方向的背景样本学习到滤波器中,利用卡尔曼滤波对目标运动状态进行估计,预测目标的运动方向。在滤波器训练时,对目标运动方向上的背景样本训练时赋予较多的权重;接着,在模型更新时引入一个新的遮挡判定指标APCE,只有当响应峰值和APCE数值分别一定比例大于各自的历史均值时,才对目标模型进行更新;最后将本文算法与当前一些主流的跟踪算法在CVPR 2013 Benchmark进行对比实验。仿真实验结果表明,本文算法的精准率和成功率分别为0. 810和0. 701,均优于其他算法,充分体现出了本文提出算法的鲁棒性。
Aiming at the problem of background information filtering too smooth when implementing equivalent weight training to context sample in context-aware correlation filter tracking algorithm,we propose an adaptive context-aware correlation filtering algorithm.And in order to solve the problem of target occlusion,we introduce a new occlusion criterion.First of all,extract background samples from the four directions of the target to learn in the filter.The target motion state is estimated by Kalman Filters and the direction of the target is predicted.During the training of the filter,more weight is given to the background sample training in the direction of the target movement.Then,a new occlusion indicator Average Peak-to correlation Energy(APCE)is introduced when the model is updated.The target model is updated only when the response peaks and APCE values are in proportional higher than their respective historical averages.Finally,the proposed algorithm is compared with some mainstream tracking algorithms in CVPR 2013 Benchmark.Simulation results show that the accuracyrate and success rate of the proposed algorithm respectively are 0.810 and 0.701,which are superior to other algorithms.The results fully reflect the robustness of the proposed algorithm.
作者
刘波
许廷发
李相民
史国凯
黄博
LIU Bo;XU Ting-fa;LI Xiang min;SHI Guo kai;HUANG Bo(School of Optics and Photonics,Laboratory for Opto-Electronic Imaging Technology,Beijing Institute of Technology,Beijing 100081,China)
出处
《中国光学》
EI
CAS
CSCD
北大核心
2019年第2期265-273,共9页
Chinese Optics
基金
国家自然科学基金重大科学仪器专项(No.61527802)~~
