摘要
运动目标跟踪作为目前计算机视觉领域的一个研究热点,存在目标旋转变形、运动模糊和背景混杂等难点.针对这些难点情况下多域卷积神经网络目标跟踪算法(MDNet)失效的问题,本文提出一种基于目标分割的多域卷积神经网络跟踪算法,旨在利用分割网络出色的目标定位能力,为MDNet网络构建一种新的网络更新方法.在跟踪过程中,通过目标分割对失效结果进行校正,重新获得目标的精确位置,再将分割获得的目标框作为样本来更新MDNet网络,有效减少样本库中正样本的背景信息干扰,提高网络的分类能力,使算法更具鲁棒性.本文所提算法在OTB50和VOT2015进行测试,与MDNet算法相比,平均跟踪精度提升3.05%,平均成功率提升2.76%.
As a research hotspot in the field of computer vision,moving object tracking technique has many difficulties such as object rotation deformation,motion blur,and background clutter.Under these difficult conditions,multi-domain convolutional neural network object tracking algorithm(MDNet)frequently appears tracking failure.This paper proposes a multi-domain convolutional neural network tracking algorithm based on object segmentation,which aims to build a new network update method for MDNet networks by taking advantage of the excellent target location capabilities of segmented network.During the tracking process,the tracking failure results are corrected through object segmentation to obtain the precise position of the object.Then update MDNet network using the target frame obtained from the segmentation as a sample,which can effectively reduce the background information interference of the positive samples in the sample database,improve the classification ability of the network,and make the algorithm more robust.The proposed algorithm was tested in OTB50 and VOT2015 dataset.Compared with MDNet algorithm,the average tracking accuracy of ours is increased by 3.05%and the average success rate is increased by 2.76%.
作者
毛琳
巩欣飞
杨大伟
张汝波
MAO Lin;GONG Xin-fei;YANG Da-wei;ZHANG Ru-bo(College of Mechanical and Electronic Engineering,Dalian Minzu University,Dalian 116600,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第5期1044-1049,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61673084)资助
辽宁省自然科学基金项目(20170540192,20180550866)资助.
关键词
目标跟踪
卷积神经网络
分割网络
网络更新
object tracking
convolutional neural network
segmentation network
network update
