QQ客服 官方微博 反馈留言

题名	基于图像局部特征的康复机器人目标识别方法研究
作者	聂海涛
学位类别	博士
答辩日期	2015-05
授予单位	中国科学院大学
导师	龙科慧
关键词	SIFT算法 图像局部特征 康复机器人 模糊控制 人脸识别 Adaboost 算法 摄像机标定
其他题名	Study of Rehabilitative Robtics Object Recognition Method Based on Local Image Feature
学位专业	机械电子工程
中文摘要	康复机器人通过视觉系统感知外部环境，将视觉信息反馈给控制系统，实现康复机器人导航定位以及机械手臂对任务目标的识别和操控，以辅助残疾人完成日常生活中的各种行为功能。康复机器人目标识别方法研究是实现机器人感知外部环境以及机器人操控的关键问题。目标特征提取与匹配是目标识别中的关键技术。提取对视角、尺度、旋转和光照具不变性的图像局部特征，对于复杂背景下目标识别的最终效果有着决定性的影响。同时，图像局部特征提取准确度直接影响康复机器人摄像机标定以及视觉系统对任务目标的姿态估计。针对康复机器人用户对象是残疾人，康复机器人目标识别的另一个主要工作是视觉系统对用户人脸信息的识别。本文根据康复机器人FRIENDⅢ研制需求，围绕基于图像局部特征的康复机器人目标识别方法关键问题展开研究，论文的主要研究工作如下： 1 研究了康复机器人目标识别的主要任务。概述了机器人视觉目标识别方法的关键问题。对比了基于全局信息和基于局部特征的目标识别图像处理算法，依次分析了典型的目标特征提取方法，特征描述符构建以及图像特征匹配策略。 2 提出了基于快速SIFT算法的目标识别方法。SIFT算法存在的主要不足是高维数的SIFT特征描述符计算复杂，造成算法实时性较差。为简化算法计算复杂程度，同时保证不损失正确匹配特征，首先构建目标图像尺度空间，提取SIFT特征点时将其按大小分类，然后扩展SIFT角度属性，由SIFT特征点子区域方向直方图计算得到4个新角度，代表特征点方向信息，最后在特征匹配时，根据SIFT特征点角度信息以及大小来限制特征点匹配范围，简化算法复杂程度，得到快速SIFT算法。实验结果表明，应用快速SIFT算法有效提高了特征匹配效率。 3 为了实现复杂背景下目标识别的稳定性，在快速SIFT算法基础上，提出了基于图像尺度因子的SIFT特征匹配方法。康复机器人FRIENDⅢ的视觉系统主要任务之一在视觉传感器发生视角变化、目标局部遮挡、旋转、以及光照变化等复杂背景下进行目标识别，但是复杂背景下能够获取的SIFT特征点数目减少，识别准确率降低。通过计算目标图像和待识别图像之间的尺度因子，在尺度因子约束条件下进行目标特征匹配，有效的保证了正确匹配数量，同时将最近邻和次近邻特征点距离限定在一个特定范围查询，减小其比值，有效地恢复被错误排除的正确匹配，保证了复杂背景下目标识别的准确率。 4 提出了一个基于扩展SIFT特征点的闭环模糊控制方法。分别提取原目标图像与待识别目标图像的极大特征值点与极小特征值点，进行SIFT特征点匹配，将每一个仿射变换与单位矩阵的差异传递给模糊控制器，用于改进匹配结果。应用Mamdani模糊控制器，采用三角模型函数和梯形模型实现模糊化，通过康复机器人目标识别实验分析，构建合理的模糊规则表，采用质心逆模糊化方法，最终实现基于扩展SIFT特征点的闭环模糊控制目标识别优化策略。 5 进行了康复机器人视觉系统目标姿态估计与摄像机标定研究。在闭环模糊控制策略的基础上，对机器人目标图像进行目标姿态估计实验，通过实验数据从图像信息中获取三维空间物体几何信息，通过计算摄像机成像几何模型参数实现康复机器人视觉系统摄像机标定。研究了基于图像的康复机器人视觉伺服方法。 6 提出了Adaboost 算法与改进SIFT方法融合的机器人用户人脸信息识别方法。康复机器人FRIENDⅢ的视觉伺服系统需要实时采集用户人脸信息，通过基于Haar-like人脸特征的Adaboost机器学习算法进行人脸检测，再通过改进SIFT算法提取人脸局部特征，实现在光照、表情、姿态发生变化情况下的人脸识别。采用SIFT算法进行人脸识别不需要对人脸图像规范化，也不需要训练样本，同时改进SIFT算法，从速度和鲁棒性方面优化算法，满足视觉系统对用户人脸识别的要求。
英文摘要	In order to assist people with disabilities to complete a variety of behaviors in daily life, rehabilitation robot perceives the external environment through vision system and sends the visual information back to its control system to realize navigation and positioning or control its mechanical arms to recognize and manipulate the object. Object recognition mothod is the key issue with regard to rehabilitation robot perceives the external environment information and manipulate. Object features extraction and matching is a key technology of object recognition. Local features extracted from the image which are invariant with respect to the viewpoint change, scale change ,rotation and illumination conditions has a decisive influence on the final results of object recognition in cluttered background. Meanwhile, the accuracy of local image features extraction directly affects camera calibration of the rehabilitate robot and its visual systems’ pose estimation for the objects. Based on the fact the user of the rehabilitate robot are the disabled people or old people, another major work for the rehabilitate robot visual system is to identify the facial information of its user. According to the research needs of rehabilitate robot FRIENDⅢ, this paper implements the research around key issues such as object recognition method based on local features of image, etc. The main research works are as follows: 1. Identify the main tasks of rehabilitate robot object recognition. Give an overview of the key issues about robot vision object recognition methods and a comparison of object recognition methods between based on global information and local features. Analyze the typical object features extraction methods, feature descriptors building and image feature matching strategy successively. 2. Propose an object recognition method based on fast SIFT algorithm. The main flaw of the SIFT algorithm is that high dimensionality of SIFT feature descriptors adds the computational complexity which leads to its poor real-time performance. In order to simply the calculation complexity without losing accurate matching features, a scale space of object image is established firstly. After the key-point detection, the SIFT features are split into two types based on its size. Then extend the SIFT angle property and four new angles are computed from the sub-region orientation histogram, which represent the orientation information of each SIFT feature. Finally, the progress of feature matching is limited in a range based on SIFT features’ angles and size, which leads to a significant simplification of the algorithm, thus a fast SIFT algorithm is obtained. The object recognition experimental results show that the fast SIFT algorithm effectively improve the efficiency of object recognition. 3. In order to achieve robust object recognition under cluttered background, a SIFT features matching method based on image scale factor is proposed according to fast SIFT algorithm. One of the main tasks of rehabilitate robot FRIENDⅢ vision system is to accomplish the mission of object recognition under cluttered conditions, such as object partial external occlusions, rotation and illumination changing. However, the number of SIFT features can be extracted is reduced which leads to lower rate of recognition accuracy. By calculating the scale factor between object image and target image, the progress of object recognition is operated under the constraint determined by the scale factor that can guarantee the number of correct matches. At the same time, the distance query between nearest-neighbor and next nearest neighbor feature point is limited to a particular range. By reducing their ratio, it can effectively restore the correct matches which are excluded by mistake to ensure target recognition accuracy under complex background. 4. A fuzzy closed loop control strategy based on extended SIFT feature is proposed. Maximum and minimum feature points of the object image and target image are extracted respectively for the use of SIFT feature points matching. The difference between each one affine transformation and unit matrix are passed to the fuzzy controller for improving the matching results. With Mamdani fuzzy controller, fuzzification is achieved by implementing triangular and trapezoidal model. A reasonable fuzzy rule table is constructed through the experimental analysis of the rehabilitate robot object recognition. By taking advantage of centroid defuzzification method, a closed loop fuzzy control object recognition optimization strategies based on extension of SIFT feature points is achieved ultimately. 5. The study of target pose estimation and camera calibration of the rehabilitate robot vision system are conducted. The closed loop fuzzy control strategy based on SIFT feature points are applied to estimate the pose of the robot target image. The geometric information of the three-dimensional spatial object is gained from the image information through the experimental data. Camera calibration of the rehabilitate robot vision system is achieved by calculating the geometric model of camera imaging. The study of rehabilitate robot image-based visual servo method is also conducted. 6. The robot user face recognition method based on Adaboost algorithm combined with improved SIFT method is proposed. The visual servo system of rehabilitate robot FRIENDⅢ requires real-time information collected form user face. The face detection is processed through Adaboost machine learning algorithm based on Haar-like facial features. Face recognition are realized under the cluttered conditions where the illumination, posture or expression are changing by using the improved SIFT algorithm to extract the local features of human face. The SIFT algorithm for face recognition has no needs to normalize face image or train samples. Meanwhile the improved SIFT algorithm most focused on optimizing algorithm speed, which can meet the real-time face recognition requirements of the visual system.
公开日期	2015-12-24
内容类型	学位论文
源URL	[http://ir.ciomp.ac.cn/handle/181722/48885]
专题	长春光学精密机械与物理研究所_中科院长春光机所知识产出
推荐引用方式 GB/T 7714	聂海涛. 基于图像局部特征的康复机器人目标识别方法研究[D]. 中国科学院大学. 2015.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

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。