17级EE转CS美MASTER求定位

274948910UID: 3763893

第三段：
除了学习相关的课程，我也从事大量科研活动，这些活动让我后来更加坚定想要继续我在我的graduate degree 和深入embeded system领域。在大二，我在浙江大学智能车实验室师从professor韩涛(Tao Han professor)从事科研。主要负责项目是设计基于视觉图像处理的自动循迹小车。我们采用以ARM cotex-M4为内核的Freescale Kinetis系列芯片为主控，在低速情况下，我们完成了智能车路径检测、速度检测、数据传输模块、电机舵机驱动模块等的设计与实现，但是在高速情况下，小车性能比较令人沮丧：我们发现由于采用大视场端焦距镜头成...像图像发生...畸变，主要有梯形...失真和桶形...失真，对图像....分析和图像...测量带来的误差，因此在高速行驶时容易造成...误判。开始我们采用常用的一些方法如：透xxx视...变换、桶形...变换 对图像...畸变....进行校..正，但是由于校正过程计算量很大，CPU占用率太高，整体系统效率降低。之后决定采用模板标定的方式，结合实际图像和采集图像，标定模板，构建拍摄模型，得到几何变换矩阵，转换成校正表写入单片机程序，同时对图像中的轨迹信息进行定向提取，极大程度的简化了图像校正的计算。最终自动循迹小车的性能大大提升，平均速度提升了1倍。当我的小车顺利完成了跑到测试，我对自己设计的系统感到满意和自豪，同时我也受到了指导老师的高度认可。

麻将小白UID: 3597343

Amazed by the power of computer, I decided to pursue Computer Architecture and Embedded System as my undergraduate learning direction, consequently I entered the electrical and information engineering departement in Zhejiang University to prepare myself with a solid foundation. During my undergraduate study, I eagerly seek for my

zsmzyUID: 3766635

a

zsmzyUID: 3766635

a

zsmzyUID: 3766635

Dear Admission Officer:

I am xxx, the manager of Processor Application Development Department at the XXX(推荐机构名). It is a great pleasure to write this recommendation letter for XXX, one of my favorite summer interns during last years’ bank loan auditing work. His earnest working attitude, excellent interpersonal skills and solid academic foundation left me a deep impression.
To begin with, his earnest working attitude struck me the most. Borrowers are required to provide numerous documents for review and archive purposes and they need to be inquired carefully within limited time period. Often times, multiple borrowers need this service at the same time. Based on my observations and past experience, most interns were not able to handle this amount of workload under such big pressure, therefore incomplete documents and auditing results were frequently seen among interns. However, to my surprise, the customers assigned to XXX barely had such problem. I found out later that his excellent work was benefited mainly from his attentive notes-taking habit. While checking his work notes, all files were carefully categorized and well documented. While recording every problem that he found, no major misconducts were ever overlooked. In fact, the work that he produced was so spotless that it left me no space for feedback or improvement suggestions. And I value this ability as one of the most important qualities that an accountant should acquire.
In terms of interpersonal and communication skills, XXX also showed extraordinary cooperation ability during the team project. Altogether 5 interns were assigned the task of bank loan audit together as a team. When XXX noticed his teammates’ growing inefficiency problem mentioned above, he pro-actively organized a group meeting before I did and shared all his personal tips and recommendations to his peers. As a result, the detrimental consequences were contained and the overall quality of the intern group was improved significantly. And of course, XXX has been recognized as the little master among the interns.
Excellent academic knowledge was also revealed. Despite the discrepancies between the Chinese accounting system that we use and the US GAAP that XXX learned at school, he quickly self-learned the key knowledge points that are critical to his daily work. More importantly, his US GAAP background provided a different angle for our work and made it extremely deep and diversified in ways I never could have imagined. I clearly remember that once we met an accountant whose company borrowed some billion RMB to build an industrial park. XXX proficiently reviewed the construction cost and gross profit balances to see whether they conformed to the gross profit recognition method that the company chose for its long-term construction contract, recorded the construction situation, whether the company foresee long-term contract losses and how that would affect their bank loan repayment.
To sum up, XXX is one of the most outstanding interns that I came to know throughout my 5 years of work experience here at XXXX. We extended a warm return offer for him, hoping that he could continue to bring positive influence to us. While when he expressed his will to study in the U.S., what I can give is only my full support. I believe, if accepted, Shuo will become a valuable addition to your education institution. Thanks for you kindly consideration.
Sincerely,


XXXXX（推荐机构）

zsmzyUID: 3766635

a

zsmzyUID: 3766635

a

zsmzyUID: 3766635

4

麻将小白UID: 3597343

Dear Sir or Madam:

I am Junjie Wang, the leader of Software Development Department in Samsung Semiconductor China Institute. This letter is to serve as my formal recommendation for Zeyang Ma. During the past 5 months, Zeyang has been an intern in our department and was in charge of two projects on Virtual Reality. As her direct project supervisor, I have a daily talk with her to track on the project progress, and I believe with such familiarity I am fully qualified to talk about what characteristic I find on her.

Firstly, I have to say Zeyang has an extraordinary learning ability and a solid programming background. In her first project of Orientation Tracking for Virtual Reality Glasses, she has to read the API code of VR glasses in order to better implement the tracking algorithm. Considering the huge volume of the code, I didn’t expect that Zeyang can locate and understand the underlying code within several days but she successfully did it. With her quick pace to comprehend the frame of the project, she completed her task much earlier than anticipated. What’s more, I can’t be more satisfied with her programming style, which is especially neat and formal, with necessary and condense commenting.

Secondly, I find Zeyang a

麻将小白UID: 3597343

aa

麻将小白UID: 3597343

aaaa

麻将小白UID: 3597343

5 匹配低层特征
目前为止，支撑相机定位问题的几何学已经被讨论了。这个问题的公式化与解决方案也已经被提出。尽管初始化一直是必须的（并且可能相当复杂），基于边缘的模型跟踪（见3.2节）与模板跟踪算法（见4.2节）可以作为跟踪方法，并且它们是完备的。对于其他方法，PnP，SLAM，点对应的单映射，低层特征提取与匹配过程是必须的。一个对文献中所有方法的综合回顾似乎超出了范畴[136]。在这一节，我们回顾在真实AR系统中被考虑的主要解决方案。

5.1 基准标记检测与定位
基准标记是发展AR应用的基本工具。实际上，它们使得同时进行目标识别与相机定位成为可能。这些标记一开始从[106]中被提出，并且延伸到了ARtoolkit[54]、ARToolkit plus[145]、Studierstube Tracker以及ARTag
[37], [38]。为了指明检测过程与基础的图像处理算法，它们的设计最终被简化了。通常它们会使用方形形状与二元颜色的结合（见图22）。更精确地说，长方形形状首先在二元化的图形中被检测，之后与长方形形状相关的相机姿态通过已知标记——四个角的3D坐标计算得到，其中使用的计算方法类似于在3.1节或4.1.3节中提到的。标记内的材质是为标记识别唯一设计的。圆形形状也常被选择以替代方形形状[86]。由于单个圆对于相机定位是不够的，多个圆被随机地[139]，圆状地[14]或者方状地[15]分布在一个平面上。
尽管关于标记的研究还在进行当中，1990年代后期发展起来的关键点匹配方法让增强现实达到了一个新的成熟程度。


5.2 关键点匹配
在前面的部分，我们提到点对应需要提前获得。这些对应一是图片中的2D点与PnP推理模型（3.1节）中的3D点的对应，二是在单应估计（4.1节）中是位于一个平面上的两个2D点间的对应。
在文献当中，1999年提出的SIFT[72]被视作2D点匹配的一个突破，之后，各种类型的关键点检测器与描述符被考虑了。2D匹配的通用框架一般考虑三个步骤：关键点提取、关键点描述与关键点匹配。首先，依据“角点”标准在图像像素中选取一系列兴趣像素。对每个选出的像素，他们的本地材质被转化为描述符（一个用于以某一特殊方式编码关键点及其邻域的向量）。最后，两张图像中提取的描述符被匹配从而确定对应关系。
当姿态估计或者单应估计被考虑时，参考模型（3D模型或图像模型）中的关键点描述符讲提前被离线计算，并存储在描述符数据库中。之后在在线过程中，关键点从各个图像中被提取，继而与描述符数据库中的描述符空间相匹配。最后，相机姿态或位移可以通过这些对应被计算出（见图16与图23）。

特征提取
从已获取的图像中，本地特征通过例如“角点”的纹理提取。理想情况下，由于相机可以在AR应用中自由移动，这些特征应该从各自的转化图像中提取。这个过程需要有很高的可重复性与实时性。因此，当前的关键点检测器通常拥有面对几何变化特征不变性，这些几何变换包括位移，旋转，仿射和大小变换。
历史上，Harris检测器[47]被广泛用作角点检测器，它根据图像块的梯度计算每一个像素的角点可能性。之后，角点可能性依据图像块的强度结构被分类为平点、边缘点和角点。SUSAN是一个选择角点像素的替代方法，区别于计算梯度，这个方法通过自我相似性[125]判断，如果一个像素与图像区块内其他像素不相像，它就被判为角点。FAST[108]沿用了SUSAN的方法，但它为了提高提取速度只考虑一个圆圈内的像素。由于FAST只计算相比于那些通过机器学习方法选出来的像素的相似性，它在计算上是非常快的。AGAST[77]相比FAST进一步改善了计算开销问题，因为它动态改变了相似性检验中的像素优化配置。
由于之前提到的关键点不拥有尺度不变性，我们可以考虑一个图像金字塔，从而关键点可以在大小改变时也被检测到。但是为了解决这个大小问题，一些基于尺度空间理论的拥有尺度不变性的检测器也被提出[70]。通常，一个线性高斯尺度空间被建立，接着该空间内的局部极值被选为关键点。其中一种尺度不变关键点检测器使用拉普拉斯高斯算子（Laplacian of Gaussian, LoG）[71]。但是为了效率问题，LoG被使用SIFT[72]中的高斯差分来近似，并且它可以进一步用GPU[123]进行加速，从而它能够被应用在增强现实中。在SURF[12]，Hessian行列式被用作另一个尺度空间运算子，并且它处理积分图像非常高效。最近，KAZE[2]使用一种非线性扩散滤波器作为非线性尺度空间，从而物体边界可以被保留，关键点提取变得更加有鲁棒性，同时，它也加速了实时检测[3]。需要注意的是，在选取之后，通常使用非最大值抑制方法，这个方法只选择一个区域内的局部角点分数极值。这是因为前面的过程中，一些冗余的关键点可能也会被提取出，这可能会造成错误的坐标对应[88]。


特征描述
下一个步骤通常是计算能够充分描述关键点与其邻域的特征向量。为了鲁棒性问题，最终的描述符需要拥有对于几何变换与光度变换的不变性。旋转不变性通常可以由旋转归一化（对提取的关键点赋旋转值）取得。













5.3 其他框架
最近，关键点对应已经被转化为分类问题[66], [95]。相对于之前提到的经典框架，一个点在仿射变换下的视图集以一个类的方式被压缩描述与处理。在运行时刻，分类技术被用于决定提取出的关键点属于哪一类。在这种方法中，一些统计学分类工具例如随机树[66]与随机蕨[95]被使用。
强化关键点间的集合约束也可以被用于简化匹配。在[138]中，它提出用几何特征而不是本地图像块进行匹配关键点。相邻关键点之间的几何关系被作为特征从而各种丰富的二元化纹理可以被检测到。
另一个有趣的方法是对无纹理对象考虑基于轮廓的方法[29], [80]。在[29]中，物体轮廓通过MSER被提取出[82]，交叉比率通过各轮廓二重切线形成的描述符被计算得知。


5.4  选择最匹配的技术
我们很难说某一方法要优于另一个。我们通常需要在稳定性、提取点数、回调、异常点比例、计算消耗等因素之间权衡。需要提出的是大部分底层对应方法都在OpenCV和VLFeat [142]中被提出。这是对于SIFT, SURF, FAST, BRIEF, ORB, MSER 等的情况。于是，在特定场景中选择效率最高的方法变得十分简单。在US注册专利的SIFT这几年被证明[72]是一个非常有效和优秀的选择（尽管计算开销比较高）。从实用性的角度看，似乎FAST是增强现实库中常被用到的算法，例如它被用在来自高通（Qualcomm）的Vuforia中或者是Zappar中。



6 结论
本次研究主要概括了实时增强现实中的相机定位问题。我们主要关注于把姿态估计的几何学部分看做对齐问题。我们也提供了一些关于这个过程中低层图像处理技术的提示。我们写这篇研究的目标是为与增强现实相关的研究者，从业者以及学生们提供指引。我们希望文章中展现的材料和例子能满足我们一开始的目标。
我们关注增强现实应用中的一个基础工具。尽管这个领域有了非常大的发展，目前还存在许多问题要被解决。五年前，追踪的可靠性与鲁棒性（对于遮挡，相机快速运动，杂乱场景等）是一个很明显的问题。现在他们被显著改善了（由于计算机视觉技术以及如IMU的传感器的引入）。除了相机定位问题，研究者们也可以考虑诸如遮挡检测与处理、动态场景、光源方向等问题。为了精确的渲染的过程，这些问题都需要被仔细考虑。
上个世纪同样也见证了许多涉足增强现实的公司与新生企业的发展。然而这个产业中很少出现一些“绝杀应用”[87]。反而，大部分被提出的系统还只是雏形。无论是在学术界还是工业界中，解决方案规模、终端用户和市场接受度显然都是潜在的提升领域。

zsmzyUID: 3766635

1

zsmzyUID: 3766635

3

zsmzyUID: 3766635

2