液晶与显示, 2018, 33 (7): 596, 网络出版: 2018-11-25   

基于卷积神经网络的响应自适应跟踪

Response adaptive tracking based on convolution neural network
作者单位
河北工业大学 控制科学与工程学院, 天津 300130
摘要
针对目标跟踪中的遮挡、旋转、快速运动、形变等问题,本文提出基于卷积神经网络的响应自适应跟踪算法。首先,通过卷积神经网络提取目标的多层卷积特征,利用粒子滤波算法获取目标的多模板响应图,自适应学习目标的期望响应; 然后通过构造目标函数的对偶形式解决多模板联合优化问题,计算多模板情况下每层卷积特征的最优滤波参数; 最后通过相关滤波算法计算多层滤波响应,通过响应加权融合的方式计算最终响应图,以此估计目标位置。本文利用OTB-2013数据集中的方法测试我们提出的算法,实验表明该算法的整体成功率和精确度为0.884和0.915。本文算法在距离准确度、成功率和平均跟踪误差方面均优于传统的相关滤波跟踪算法,有一定研究价值。
Abstract
In order to solve the problem of the occlusion, rotation, fast motion, deformation in target tracking, the paper proposes the response adaptive tracking algorithm based on convolution neural network. First, we extract multi-layer convolutional features of target by using convolution neural network, and gain the multi-template response of the target exploiting particle filter algorithm to adaptively learn the objectives of the expected response. Then, the dual form of the objective function is constructed to solve the multi-template joint optimization problem in order to calculate the optimal filtering parameters of each-layer convolutional features in the multi-template case. Finally, we calculate the multi-layer response by utilizing correlation filter algorithm and calculate the final response map by using the weighted fusion method, and then the proposed algorithm estimates the target position by employing the final response map. In this paper, we use the method of OTB-2013 data set to test the algorithm, experimental results show that the overall success rate and accuracy of the algorithm are 0.884 and 0.915,repectively. The algorithm is better than the traditional correlation filter tracking algorithm in distance precision, success rate and average tracking error, so it has a certain research value.

李勇, 杨德东, 毛宁, 李雪晴. 基于卷积神经网络的响应自适应跟踪[J]. 液晶与显示, 2018, 33(7): 596. LI Yong, YANG De-dong, MAO Ning, LI Xue-qing. Response adaptive tracking based on convolution neural network[J]. Chinese Journal of Liquid Crystals and Displays, 2018, 33(7): 596.

本文已被 2 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!