基于深度学习的人体姿态估计方法综述

基于深度学习的人体姿态估计方法旨在通过构建合适的神经网络，直接从二维的图像特征中回归出人体姿态信息。主要按照2D人体姿态估计到3D人体姿态估计的顺序，并从单人检测与多人检测、稀疏的关节点检测与密集的模型构建等方面，对近年来基于深度学习的人体姿态估计方法进行系统介绍，从而初步了解如何通过深度学习的方法得到人体姿态的各个要素，包括肢体部件的相对朝向和比例尺度、骨骼关节点的位置坐标和连接关系，甚至更为复杂的人体蒙皮模型信息。最后,对当前研究面临的挑战以及未来的热点动向进行概述，清晰地呈现出该领域的发展脉络。     

一种新型电子显示屏的研制

本文介绍一种不同于磁性翻牌显示屏和ＬＥＤ显示屏的新型电子显示屏，采用非发光体，针对显示屏彩色和动画图象对象素的要求，讨论了显示屏象素的多面色体变换的电路设计及软件实现。     

非透视3点问题的解析解

由于在机器人导航等领域新的图像装置不符合透视规则,无法运用传统的透视位置估计方法,采用吴氏零点分解方法,给出非透视3点问题的解析解.利用新图像装置的几何条件,建立了非透视3点方程系统.     

Extending the color gamut and dynamic range of an sRGB image using a residual image

Image sources, such as digital camera captures and photographic negatives, typically have more information than can be reproduced on a photographic print or a video display. The information that is lost during the tone/color rendering process relates to both the extended dynamic range and color gamut of the original scene. In conventional photographic systems, most of this additional information is archived on the photographic negative and can be accessed by adjusting the way the negative is printed. However, most digital imaging systems have traditionally archived only a rendered video RGB image. As a result, it is not possible to make the same sorts of image manipulations that historically have been possible with conventional photographic systems. This suggests that there would be an advantage to storing images using an extended dynamic range/color gamut color encoding. However, because of file compatibility issues, digital imaging systems that store images using color encoding other than a standard video RGB representation (e.g., sRGB) would be significantly disadvantaged in the marketplace. In this article, we describe a solution that has been developed to maintain compatibility with existing file formats and software applications, while simultaneously retaining the extended dynamic range and color gamut information associated with the original scenes. With this approach, the input raw digital camera image or film scan is first transformed to the scene‐referred ERIMM RGB color encoding. Next, a rendered sRGB image is formed in the usual way and stored in a conventional image file (e.g., a standard JPEG file). A residual image representing the difference between the original extended dynamic range image and the final rendered image is formed and stored in the image file using proprietary metadata tags. This provides a mechanism for archiving the extended dynamic range/color gamut information, which is normally discarded during the rendering process, without sacrificing interoperability. Appropriately enabled applications can decode the residual image metadata and use it to reconstruct the ERIMM RGB image, whereas applications that are not aware of the metadata will ignore it and only have access to the sRGB image. The residual image is formed such that it will have negligible pixel values for those portions of the image that lie within the sRGB gamut, and will therefore be highly compressible. Tests on a population of 950 real customer images have demonstrated that the extended dynamic range scene information can be stored with an average file size overhead of about 8% compared to the sRGB images alone. © 2003 Wiley Periodicals, Inc. Col Res Appl, 28, 251–266, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/col.10160     

Chromaticity contour map of the RGB cube: A simple algorithm

A simple algorithm (amenable to spreadsheets) is described for plotting a luminance contour map of the color gamut of an additive‐primary display given the chromaticities of its primaries and screen white. Each contour of the plot represents (in either r,g or x,y coordinates) the polygon of intersection of the RGB cube with a plane of constant Y. Each vertex on a contour is an intersection of 1 of the 12 edges of the RGB cube with the constant‐Y plane. Feasibility of an edge as a vertex of the polygon is tested through comparison of the computed r,g coordinates with functions of the luminosity coefficients of the primaries. By ordering the RGB‐cube edges in a particular way prior to scanning them for feasibility, one is guaranteed to retrieve the vertices of any constant‐Y polygon in convex order. This fact facilitates plotting of the polygons.     

A Multi-View Face Recognition System

In many automatic face recognition systems,posture constraining is a key factor preventing them from application.In this paper a series of strategies will be described to achieve a system which enables face recognition under varying pose.These approaches include the multi-view face modeling,the threschold image based face feature detection,the affine transformation based face posture normalization and the template matching based face identification.Combining all of these strategies,a face recognition system with the pose invariance is designed successfully,Using a 75MHZ Pentium PC and with a database of 75 individuals,15 images for each person,and 225 test images with various postures,a very good recognition rate of 96.89% is obtained.     

Robot Pose Estimation in Unknown Environments by Matching 2D Range Scans

A mobile robot exploring an unknown environment has no absolute frame of reference for its position, other than features it detects through its sensors. Using distinguishable landmarks is one possible approach, but it requires solving the object recognition problem. In particular, when the robot uses two-dimensional laser range scans for localization, it is difficult to accurately detect and localize landmarks in the environment (such as corners and occlusions) from the range scans.In this paper, we develop two new iterative algorithms to register a range scan to a previous scan so as to compute relative robot positions in an unknown environment, that avoid the above problems. The first algorithm is based on matching data points with tangent directions in two scans and minimizing a distance function in order to solve the displacement between the scans. The second algorithm establishes correspondences between points in the two scans and then solves the point-to-point least-squares problem to compute the relative pose of the two scans. Our methods work in curved environments and can handle partial occlusions by rejecting outliers.     

运动捕捉数据在Jack平台上的重定向技术研究

提出了一种BVH格式运动捕捉数据驱动Jack三维骨架模型产生人体运动效果的方法。将Peabody结构的Jack虚拟人模型简化成能够映射BVH数据的树状骨骼结构,使用欧拉角旋转方程建立运动捕捉数据与Jack角色模型的关节数据映射公式,最后在Jack平台上用Python等脚语言进行了编程实现。为在Jack平台中大规模重用运动捕捉数据提供了条件。     

局部特征表征的6D位姿估计算法

为解决有纹理模型在遮挡条件下6D位姿估计精确度不高的问题,提出了一种局部特征表征的端到端6D位姿估计算法。首先为了得到准确的定位信息,提出了一个空间—坐标注意力机制（spatial and coordinate attention）,通过在YOLOv5网络中加入空间—坐标注意力机制和加权双向特征金字塔网络（bidirectional feature pyramid network）,YOLOv5-CBE算法的精确度（precision）、召回率（recall）、平均精度均值（mAP@0.5）分别提升了3.6%、2.8%、2.5%,局部特征中心点坐标误差最高提升了25%;然后用 YOLOv5-CBE算法检测局部特征关键点,结合3D Harris关键点通过奇异值分解法（singular value decomposition）计算模型的6D位姿,最高遮挡70%的情况下仍然可以保证二维重投影精度（2D reprojection accuracy）和ADD度量精度（ADD accuracy）在95%以上,具有较强的鲁棒性。     

3D human pose estimation in motion based on multi-stage regression

3D human pose estimation in motion is a hot research direction in the field of computer vision. However, the performance of the algorithm is affected by the complexity of 3D spatial information, self-occlusion of human body, mapping uncertainty and other problems. In this paper, we propose a 3D human joint localization method based on multi-stage regression depth network and 2D to 3D point mapping algorithm. First of all, we use a single RGB image as the input, through the introduction of heatmap and multi-stage regression to constantly optimize the coordinates of human joint points. Then we input the 2D joint points into the mapping network for calculation, and get the coordinates of 3D human body joint points, and then to complete the 3D human body pose estimation task. The MPJPE of the algorithm in Human3.6 M dataset is 40.7. The evaluation of dataset shows that our method has obvious advantages.