CASIA OpenIR  > 毕业生  > 硕士学位论文
基于视频序列的火车全景成像技术研究
胡锦高
学位类型工程硕士
导师杨一平
2017-06
学位授予单位中国科学院大学
学位授予地点北京
关键词火车 视频序列 全景成像 图像拼接
摘要
图像拼接是计算机视觉中最活跃的研究领域之一,其应用包括航拍遥感图
像拼接、医学成像、安防监控、三维重建等。通过图像拼接的方法获取火车车
身的高清全景图像,可以为基于计算机视觉的火车车速测量、车型和车号识别
等技术提供便利,进而提高铁道交通管理的自动化程度和智能化水平。在对高
速运动的火车目标成像时,由于目标尺寸较大,形状极狭长,同时处于运动状
态,而相机成像视野有限,通常一帧图像只能拍摄到火车的局部,很难一次性
获取其高清全貌信息。因此,分析连续的包含火车局部信息的视频序列,通过
图像拼接还原火车的全景信息,具有重要的研究意义和应用价值。
本文设计并且实现了一套火车全景成像系统,包括硬件系统和软件系统。
系统首先采集连续的包含火车局部信息的高清视频序列,然后基于视频序列
进行火车全景成像。在视频序列的配准阶段,本文分别研究了金字塔模板匹配
法、KLT与DC Mean-Shift算法,两种方法在配准精度、配准速度以及鲁棒性
等方面可以相互补充。
本文的主要工作包括:
1. 设计并实现了火车全景成像系统,包括硬件系统和软件系统。硬件系统
基于英伟达(NVIDIA)公司的Jetson TK1开发板,采用模块化设计,其核心
处理单元为Tegra K1,同时集成了视觉传感模块、调校模块、交互模块和支撑
模块。软件系统包括图像预处理、视频序列配准、前景帧提取和图像融合等模
块。系统可以独立实现户外移动条件下的火车全景成像。
2. 实现了一种基于金字塔模板匹配的火车全景成像方法。该方法针对传统
模板匹配计算量大的问题,采用了图像金字塔结构,首先在图像的低分辨率尺
度上配准,根据配准结果进一步在高分辨率尺度配准,然后修正误匹配点、提
取前景帧并进行图像融合。实验结果表明,算法在不损失精度的同时效率明显
有所提高。
3. 针对视频序列配准问题,提出了一种基于KLT与DC Mean-Shift的算法。
该算法对每一帧图像提取一定数目的特征点并计算这些特征点在连续两帧间的
光流,然后对特征点的光流结果进行DC Mean-Shift分析。DC Mean-Shift算法
ii 基于视频序列的火车全景成像技术研究
利用了火车运动方向这一先验知识,实验结果表明,相比于Mean-Shift其收敛
速度更快,正确率更高。
其他摘要
Image mosaic is one of the most active research topics in the field of computer
vision and its application includes aerial and satellite image mosaic, medical
imaging, security monitoring and 3D reconstruction. High-definition panoramic
images of trains by image mosaic can provide convenience for computer visionbased
train speed measuring, type and carriage number recognision which can
improve the automation and intelligence level of Ministry of Railways. A train
panoramic image is hard to access once due to its big size, extremely long and
narrow shape. Instead, a frame can only cover a small part of a train for camera’s
limmitted imaging scope. As a result, the research on train panoramic imaging
based on video sequence is of great significance and applications.
In this paper, we design and implement a train panoramic imaging system
which consists of hardware system and software system. The system first records
a high definition video sequence of a train and then do the panoramic imaging. In
the stage of sequence registration, we use image pyramid-based template matching
and direction constrained KLT method respectively. The two methods can
complement each other in registration accuracy, time and robustness.
The main contributions of our work are listed as follows:
1. We design and implement a train panoramic imaging system which consists
of hardware system and software system. The hardware system is based on
a NVIDIA Jetson TK1 board and constructed by modules. The core processiong
unit is Tegra K1 and visual sensing device, adjusting device, input-output device
and support device are integrated. The software system consists of image
preprocessing, video sequence registration, foreground frame extracting and image
fusion. The system can do train panoramic imaging in outdoor environment
independently.
2. We implement a image pyramid-based template matching method. It
takes down the intense compution burden of traditional template matching method
by using a image pyramid. We first do the template matching in low resolution
and then in high resolution, next, we correct the wrong matching results, extract
foreground frames from background frames and finally do the stitching.
Experiment results show that our method works efficiently with little accuracy
loss.
3. We propose a KLT and DC Mean-Shift based method for sequence registration.
We first compute the optical flow of some feature points and then use
direction constrained Mean-Shift(DC Mean-Shift) to do an analysis of the optical
flow results, compute a final value as the displacement of the train between two
successive frames. DC Mean-Shift takes advantage of priori knowledge of train
moving direction and converges faster and is more accuracy than Mean-Shift.
Keywords: Train, Video Sequence, Panoramic Imaging, Image Mosaic
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14731
专题毕业生_硕士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
胡锦高. 基于视频序列的火车全景成像技术研究[D]. 北京. 中国科学院大学,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
template.pdf(13369KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[胡锦高]的文章
百度学术
百度学术中相似的文章
[胡锦高]的文章
必应学术
必应学术中相似的文章
[胡锦高]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。