基于深度学习的二维人体姿态估计综述

人体姿态估计是计算机视觉领域的一个基础且具有挑战的任务,人体姿态估计对于描述人体姿态、描述人体行为等至关重要,是行为识别、行为检测等计算机视觉任务的基础.近年来,随着深度学习的发展,基于深度学习的人体姿态估计算法展现出了极其优异的效果.从单人人体姿态估计、自顶向下的多人人体姿态估计和自底向上的多人人体姿态估计这3种主流的人体姿态估计方式,介绍近年来基于深度学习的二维人体姿态估计算法的发展,并讨论目前二维人体姿态估计所面临的困难和挑战.最后,对人体姿态估计未来的发展做出展望.     

机械臂绝对定位精度测量

提出了用激光跟踪仪标定机械臂的D-H参数、测量机械臂绝对位姿以及对机械臂的绝对定位精度进行分析的方法;用激光跟踪仪测量机械臂各个关节单独运动时得到的一系列离散点,就可确定机械臂各个关节的轴线,由此建立机械臂的D-H坐标系,并对D-H参数进行标定;然后,给出了由6D激光头位姿确定机械臂末端位姿的方法;最后,推出了由测量位姿值与命令位姿值相比较,得到机械臂绝对定位的位置和姿态偏差的方法;这些方法可以有效、迅速地完成对机械臂绝对定位精度的测量.     

基于单视图的多姿态人脸识别算法

针对基于多视图的多姿态人脸识别方法的缺陷，即需要对每个人脸拍摄多个视图为前提条件，提出了基于单视图的多姿态人脸识别技术，首先基于二元高次多项式函数最小二乘拟合方法由单视图通过变形生成多姿态人脸图像，然后基于该单视图和生成的多姿态图像进行多姿态人脸识别。实验结果表明该文算法识别的正确率远高于经典算法。     

A comparative study of the robustness of two pose estimation techniques

The paper presents an analysis of the stability of pose estimation. Stability is defined as sensitivity of the pose parameters towards noise in image features used for estimating pose. The specific emphasis of the analysis is on determining {how the stability varies with viewpoint} relative to an object and to understand the relationships between object geometry, viewpoint, and pose stability. Two pose estimation techniques are investigated. One uses a numerical scheme for finding pose parameters; the other is based on closed form solutions. Both are “pose from trihedral vertices” techniques, which provide the rotation part of object pose based on orientations of three edge segments. The analysis is based on generalized sensitivity analysis propagating the uncertainty in edge segment orientations to the resulting effect on the pose parameters. It is shown that there is a precomputable, generic relationship between viewpoint and pose stability, and that there is a drastic difference in stability over the range of viewpoints. This viewpoint variation is shared by the two investigated techniques. Additionally, the paper offers an explicit way to determine the most robust viewpoints directly for any given vertex model. Experiments on real images show that the results of the work can be used to compute the variance in pose parameters for any given pose. For the predicted {instable} viewpoints the variance in pose parameters is on the order of 20 (degrees squared), whereas the variance for robust viewpoints is on the order of 0.05 (degrees squared), i.e., two orders of magnitude difference.     

基于改进头姿估计方法的机器人轮椅交互控制

针对现有迭代最近点（ICP）头姿估计算法存在迭代次数偏多且易陷于局部最优、而随机森林（RF）头姿估计算法准确性和稳定性不高的问题,提出一种新的头姿估计改进方法,并基于该改进方法构建机器人轮椅实时交互控制接口．首先,分析现有迭代最近点头姿算法与随机森林头姿算法在准确性、实时性及稳定性方面存在的问题,并提出一种新的基于随机森林与迭代最近点算法融合的头姿估计改进方法;其次,为实现头姿估计到机器人轮椅交互控制的无缝连接,建立基于传统机器人轮椅操纵杆的头部姿态运动空间映射;最后,在基于标准头姿数据库分析改进头姿估计方法性能的基础上,构建机器人轮椅实验平台并规划运动轨迹,以进一步验证基于改进头姿估计方法的人机交互接口在机器人轮椅实时控制方面的有效性．实验结果表明,改进后的头姿估计方法较传统迭代最近点算法减少了迭代次数且避免了陷于局部最优,在仅增加少量运算时间的基础上,其准确性和稳定性都优于传统随机森林算法;同时,基于改进头姿估计方法的人机交互接口亦能实时平稳地控制机器人轮椅沿既定的轨迹运动．     

消除光照和姿态变化影响的人脸识别

光照和姿态变化带来的影响是自动人脸识别的两个主要瓶颈问题。提出了消除这两方面影响的处理方法：首先对训练集里的图像应用灰度归一化处理,降低对光照强度的敏感度;然后进行姿态估计,并用特征脸方法计算不同姿态的特征子空间,最后提出了“姿态权重PWV（Pose’s Weight Value）”这一概念,据此设计了加权的最小距离分类器WMDC（Weighted Minimum Distance Classifier）,分配不同姿态权重消除姿态变化影响。在FERET和Yale B数据库上的实验结果表明,此方法能在很大程度上提高人脸光照和姿态改变时的识别率。     

着装场景下双分支网络的人体姿态估计

目的 人体姿态估计旨在识别和定位不同场景图像中的人体关节点并优化关节点定位精度。针对由于服装款式多样、背景干扰和着装姿态多变导致人体姿态估计精度较低的问题,本文以着装场景下时尚街拍图像为例,提出一种着装场景下双分支网络的人体姿态估计方法。方法 对输入图像进行人体检测,得到着装人体区域并分别输入姿态表示分支和着装部位分割分支。姿态表示分支通过在堆叠沙漏网络基础上增加多尺度损失和特征融合输出关节点得分图,解决服装款式多样以及复杂背景对关节点特征提取干扰问题,并基于姿态聚类定义姿态类别损失函数,解决着装姿态视角多变问题;着装部位分割分支通过连接残差网络的浅层特征与深层特征进行特征融合得到着装部位得分图。然后使用着装部位分割结果约束人体关节点定位,解决服装对关节点遮挡问题。最后通过姿态优化得到最终的人体姿态估计结果。结果 在构建的着装图像数据集上验证了本文方法。实验结果表明,姿态表示分支有效提高了人体关节点定位准确率,着装部位分割分支能有效避免着装场景中人体关节点误定位。在结合着装部位分割优化后,人体姿态估计精度提高至92.5%。结论 本文提出的人体姿态估计方法能够有效提高着装场景下的人体姿态...     

A Two-Layer Framework for Piecewise Linear Manifold-Based Head Pose Estimation

Fine-grain head pose estimation from imagery is an essential operation for many human-centered systems, including pose independent face recognition and human-computer interaction (HCI) systems. It is only recently that estimation systems have evolved past coarse level classification of pose and concentrated on fine-grain estimation. In particular, the state of the art of such systems consists of nonlinear manifold embedding techniques that capture the intrinsic relationship of a pose varying face dataset. The success of these solutions can be attributed to the acknowledgment that image variation corresponding to pose change is nonlinear in nature. Yet, the algorithms are limited by the complexity of embedding functions that describe the relationship. We present a pose estimation framework that seeks to describe the global nonlinear relationship in terms of localized linear functions. A two layer system (coarse/fine) is formulated on the assumptions that coarse pose estimation can be performed adequately using supervised linear methods, and fine pose estimation can be achieved using linear regressive functions if the scope of the pose manifold is limited. A pose estimation system is implemented utilizing simple linear subspace methods and oriented Gabor and phase congruency features. The framework is tested using widely accepted pose-varying face databases (FacePix(30) and Pointing’04) and shown to perform fine head pose estimation with competitive accuracy when compared with state of the art nonlinear manifold methods.     

3D object pose form clustering with multiple views

A candidate pose algorithm is described which computes object pose from an assumed correspondence between a pair of 2D image points and a pair of 3D model points. By computing many pose candidates actual object pose can usually be determined by detecting a cluster in the space of all candidates. Cluster space can receive candidate pose parameters from independent computations in different camera views. It is shown that use of of geometric constraint can be sufficient for reliable pose detection, but use of other knowledge, such as edge presence and type, can be easily added for increased efficiency.     

姿态特征与深度特征在图像动作识别中的混合应用

人体姿态是动作识别的重要语义线索,而CNN能够从图像中提取有很强判别能力的深度特征,本文从图像局部区域提取姿态特征,从整体图像中提取深度特征,探索两者在动作识别中的互补作用.首先介绍了一种姿态表示方法,每个肢体部件的姿态由描述该部件姿态的一组Poselet检测得分表示.为了抑制检测错误,设计了基于部件的模型作为检测上下文.为了从数量有限的数据集中训练CNN网络,本文使用了预训练和精细调节的方法.在两个数据集中的实验表明,本文介绍的姿态特征与深度特征混合使用,动作识别性能得到了极大提升.     

Adaptive end-effector pose control for tomato harvesting robots

This paper investigates the development of a tomato-harvesting robot operating on a plant factory and primarily studies the reachable pose of tomatoes in the nondexterous workspace of manipulator. The end-effector can only reach the tomatoes with reachable poses when the tomatoes are within the nondexterous workspace. If the grasping pose is not reachable, it will lead to grasping failure. An adaptive end-effector pose control method based on a genetic algorithm (GA) is proposed to find a reachable pose. The inverse kinematic solution based on analysis method of the manipulator is analyzed and the objective function of whether the manipulator has a solution or not is obtained. The grasping pose is set as an individual owing to the position of the tomatoes is fixed and the grasping pose is variable. The GA is used to solve until a pose that can make the inverse kinematics have a solution is generated. This pose is the reachable grasping pose of the tomato at this position. The quintic interpolation polynomial is used to plan the trajectory to avoid damage to tomatoes owing to fast approaching speed and a distance based background filtering method is proposed. Experiments were performed to verify the effectiveness of the proposed method. The radius of the workspace of the UR3e manipulator with the end-effector increased from 550 to 800 mm and the grasping range expanded by 208%. The harvesting success rate using the adaptive end-effector pose control method and trajectory planning method was 88%. The cycle of harvesting a tomato was 20 s. The experimental results indicated that the proposed tomato-recognition and end-effector pose control method are feasible and effective.     

基于三维模型的人脸姿态估计

提出了一种基于三维模型的人脸姿态估计方法。首先根据人脸特征点重建出稀疏的三维人脸模型,然后基于三维模型采用线性回归的方法对人脸姿态进行初步估计,确定姿态范围,再对估计结果进行修正,从而对人脸姿态进行精确估计。实验表明,该方法具有较好的估计效果,提高了姿态估计精度。     

未标定相机的高精度位姿估计方法

针对未标定相机的位姿估计问题,提出了一种焦距和位姿同时迭代的高精度位姿估计算法。现有的未标定相机的位姿估计算法是焦距和相机位姿单独求解,焦距估计精度较差。提出的算法首先通过现有算法得到相机焦距和位姿的初始参数;然后在正交迭代的基础上推导了焦距和位姿最小化函数,将焦距和位姿同时作为初始值进行迭代计算;最后得到高精度的焦距和位姿参数。仿真实验表明提出的算法在点数为10,噪声标准差为2的情况下,角度相对误差小于1%,平移相对误差小于4%,焦距相对误差小于3%;真实实验表明提出的算法与棋盘标定方法的精度相当。与现有算法相比,能够对未标定相机进行高精度的焦距和位姿估计。     

基于目标特征的SAR图像车辆目标的方位角联合估计*

针对SAR图像地面车辆目标方位角估计精度不高的问题,尤其是0度和180度估计误差大的问题,提出了一种地面SAR图像目标方位角联合估计方法。首先,分析了地面车辆目标在不同角度的成像特点和典型目标方位角估计方法的优缺点,然后,通过判断当前目标成像所具有的特点,利用目标阴影特征与目标轮廓特征,并结合目标主轴提取方法和Hough变换方法对SAR图像目标方位角进行联合估计,最后利用MSTAR目标切片数据对该方法进行了验证实验,绝对误差在5º范围内的准确估计率都在89%以上,目标误差均值都在4º以内。实验结果表明该算法的方位角估计精度比较高,算法是有效性的和可行的。     

Solution of the Simultaneous Pose and Correspondence Problem Using Gaussian Error Model

The use of hypothesis verification is recurrent in the model-based recognition literature. Verification consists in measuring how many model features transformed by a pose coincide with some image features. When data involved in the computation of the pose are noisy, the pose is inaccurate and difficult to verify, especially when the objects are partially occluded. To address this problem, the noise in image features is modeled by a Gaussian distribution. A probabilistic framework allows the evaluation of the probability of a matching, knowing that the pose belongs to a rectangular volume of the pose space. It involves quadratic programming, if the transformation is affine. This matching probability is used in an algorithm computing the best pose. It consists in a recursive multiresolution exploration of the pose space, discarding outliers in the match data while the search is progressing. Numerous experimental results are described. They consist of 2D and 3D recognition experiments using the proposed algorithm.     

Car model recognition by utilizing symmetric property to overcome severe pose variation

This paper presents a mirror morphing scheme to deal with the challenging pose variation problem in car model recognition. Conventionally, researchers adopt pose estimation techniques to overcome the pose problem, whereas it is difficult to obtain very accurate pose estimation. Moreover, slight deviation in pose estimation degrades the recognition performance dramatically. The mirror morphing technique utilizes the symmetric property of cars to normalize car images of any orientation into a typical view. Therefore, the pose error and center bias can be eliminated and satisfactory recognition performance can be obtained. To support mirror morphing, active shape model (ASM) is used to acquire car shape information. An effective pose and center estimation approach is also proposed to provide a good initialization for ASM. In experiments, our proposed car model recognition system can achieve very high recognition rate (>95%) with very low probability of false alarm even when it is dealing with the severe pose problem in the cases of cars with similar shape and color.     

Pose alignment of aircraft structures with distance sensors and CCD cameras

A precise transshipment system is developed for automatic transporting material between an automated guided vehicle (AGV) and a load transfer station. In order to implement the alignment of the fixture base on board the AGV with that of the station, it's necessary to measure the pose of the AGV with respect to the station. The pose measurement system combined four distance sensors with two CCD cameras, with the distance sensors to measure the yaw, pitch angle and longitudinal deviation, and the cameras to determine the roll angle and the position deviation in the reference plane. A 6-degree of freedom (DOF) pose alignment system based on 3-DOF positioners is used to correct the pose deviation of the current pose with respect to the goal pose, lowering the demand on locating precision of AGV on the ground. The method to calibrate the whole measurement system is elaborately stated. The transshipment experiments have been conducted and the results show that the pose alignment system integrated with the proposed multi-sensor pose measurement system can meet the accuracy and repeatability required in industrial application.     

Instance segmentation based 6D pose estimation of industrial objects using point clouds for robotic bin-picking

3D object pose estimation for robotic grasping and manipulation is a crucial task in the manufacturing industry. In cluttered and occluded scenes, the 6D pose estimation of the low-textured or textureless industrial object is a challenging problem due to the lack of color information. Thus, point cloud that is hardly affected by the lighting conditions is gaining popularity as an alternative solution for pose estimation. This article proposes a deep learning-based pose estimation using point cloud as input, which consists of instance segmentation and instance point cloud pose estimation. The instance segmentation divides the scene point cloud into multiple instance point clouds, and each instance point cloud pose is accurately predicted by fusing the depth and normal feature maps. In order to reduce the time consumption of the dataset acquisition and annotation, a physically-simulated engine is constructed to generate the synthetic dataset. Finally, several experiments are conducted on the public, synthetic and real datasets to verify the effectiveness of the pose estimation network. The experimental results show that the point cloud based pose estimation network can effectively and robustly predict the poses of objects in cluttered and occluded scenes.     

基于退火粒子群优化的单目视频人体姿态分析方法

提出一种基于退火粒子群优化(Simulated annealing particle swarm optimism, SAPSO)的单目视频人体姿态分析方法. 该方法具有以下特点: 首先, 利用运动捕获数据采用主成分分析方法(Principle component analysis, PCA)得到更能反映人体运动本质的姿态紧致空间, 并在此低维空间中进行姿态分析, 提高了姿态分析的准确性和效率; 其次, 将粒子群优化应用到姿态分析中, 并提出退火粒子群优化姿态分析方法, 该方法具有良好的收敛性和全局最优能力; 再次, 基于退火粒子群优化姿态分析方法, 实现了基于单目视频的人体姿态估计和跟踪. 实验结果表明, 本文方法不仅具有良好的计算效率, 同时具有良好的收敛性和全局搜索能力, 能准确分析单目视频中的人体姿态.     

Pose partitioning for multi‐resolution segmentation of arbitrary mesh animations

We present a complete approach to efficiently deriving a varying level‐of‐detail segmentation of arbitrary animated objects. An over‐segmentation is built by combining sets of initial segments computed for each input pose, followed by a fast progressive simplification which aims at preserving rigid segments. The final segmentation result can be efficiently adjusted for cases where pose editing is performed or new poses are added at arbitrary positions in the mesh animation sequence. A smooth view of pose‐to‐pose segmentation transitions is offered by merging the partitioning of the current pose with that of the next pose. A perceptually friendly visualization scheme is also introduced for propagating segment colors between consecutive poses. We report on the efficiency and quality of our framework as compared to previous methods under a variety of skeletal and highly deformable mesh animations.