红外与RGB信息的融合跟踪研究

摘要摘要红外与可见光信息的融合跟踪兼顾了红外与可见光设备的多源信息，可提高目标跟踪技术的性能，为国防军事、安防监控、遥感检测、资源探测等领域提供了一种有效技术。图像融合跟踪主要有图像融合和目标跟踪两个流程。依据融合层次，图像融合跟踪可被划分为像素级、特征级和决策级三大类。本文主要讨论了像素级图像融合跟踪。基于深度学习技术，图像融合跟踪可分为基于传统方法的图像融合跟踪算法和基于深度学习的图像融合跟踪算法。本文首先实现了基于TIF(Two-scale Image Fusion,TF)和均值偏移(MeanShift,MS)的目标跟踪方法。然后，本文又实现了基于卷积神经网络(Convolutional NeuralNetworks,CNN)和时空正则化的相关滤波(STRCF)的目标跟踪算法。基于此，本文给出了几种融合跟踪方法的客观评估指标（精确度和成功率）和主观视觉效果对比，并给出了对比分析。关键词：红外与可见光图像：融合跟踪：信息融合：目标跟踪：深度学习AbstractABSTRACTThe fusion and tracking of infrared and visible information takes into account themulti-source information of infrared and visible equipment,which can improve the performanceof target tracking technology,and provides an effective technology for the fields of nationaldefense and military,security monitoring,remote sensing detection,resource detection and otherfields.Image fusion tracking mainly has two processes:image fusion and target tracking.According to the fusion level,the image fusion tracking can be divided into three categories:pixel level,feature level and decision level.This paper focuses on pixel-level image fusiontracking.Based on deep learning technology,image fusion tracking can be divided into imagefusion tracking algorithm based on traditional methods and image fusion tracking algorithmbased on deep learning.In this paper,the target tracking method based on TIF (Two-scale ImageFusion,TIF)and mean shift (MeanShift,MS)is first implemented.Then,this paper implementsthe target tracking algorithm based on convolutional neural network (Convolutional NeuralNetworks,CNN)and spatial-temporal regularized correlation filters (STRCF).Based on this,thispaper gives several objective evaluation indicators (precision rate and success rate)andsubjective visual effect comparison,and gives a comparative analysis.Keywords:infrared and visible images;fusion tracking;information fusion;target tracking;deeo learning目录目录第1章绪论…1.1研究背景和意义.·········1.2国内外研究现状.·····…··1.2.1图像融合研究现状.11.2.2目标跟踪研究现状.·EEEEEEEE21.3本文主要研究成果.·.··31.4本文内容安排.·.·.····。。。。。。。。。。第2章传统信息融合跟踪……42.1T吓算法实现图像融合.········2.1.1双尺度图像分解......E.EE42.1.2显著性检测.52.1.3计算权重图…62.1.4图像融合.2.1.5图像重建.12.1.6彩色图融合.72.2 MEANSHIFT实现目标跟踪....，，，...，，，..72.2.1目标模型和候选模型描述.....822.2相似性度量…2.2.3迭代过程.……9第3章基于深度学习的R-RGB信息融合跟踪…。103.1CNN实现图像融合.··········103.1.1基于CNN生成权重图.·。103.1.2金字塔分解.113.1.3系数融合…113.1.4拉普拉斯金字塔重建...123.2 STRCF实现目标跟踪....，...，，，..123.2.1相关滤波.…。…123.2.2 MOSSE实现目标跟踪..·133.2.3DC下实现目标跟踪.....143.2.4 STRCF实现目标跟踪.·.·.15第4章实验结果…1741数据集.·····…174.2评估指标。。。。。。···。·····174.3定性分析.·17