【作者】 李晓雨；

【导师】 赵毅； 王红波；

【作者基本信息】 吉林大学 ， 集成电路工程（专业学位）， 2019， 硕士

【摘要】 机器人技术由最初的工业领域逐渐过渡至生活的各个方面,为广大人民群众的生活提供了极大便利。使机器人具有自主定位与环境建立地图的能力,是提高机器人智能化程度的一大关键问题。SLAM(Simultaneous Localization and Mapping)即时定位与建图算法,正是试图从机器人视觉图像方向解决这一问题。而该算法中机器人由于长时间运行往往会产生漂移误差,因而设计了回环检测系统判断机器人是否回到了原有的路径经过的地点,以达到消除漂移误差的目的。传统的回环检测系统使用SIFT(Scale-invariant feature transform)算法进行特征提取,虽然有着优良的精确性,但因其计算量过大,且描述子占用内存过多等缺点,导致整个算法无法应用于实时检测过程。本文针对特征提取算法与描述子进行了改进,并且针对回环检测过程中不合理的关键帧查询逻辑进行了调整,继而创新性的将机器学习算法引入了回环检测的数据库查询与匹配阶段,最终通过对所有改进点进行实验,证明了改进算法的优良性。首先,本文对Kinect摄像机进行了基本介绍并对实验所使用的相机进行了标定。其次,文章对图像的特征提取算法进行了基本介绍,引出了目前常用的较为成熟的SIFT算法,阐述了SIFT算法的步骤以及基本原理。进而,本文针对整个回环检测系统的检测流程即图片特征描述子提取、特征匹配以及匹配分组验证流程进行了介绍。通过对K-means算法的原理分析引出了用于特征匹配分组的词袋模型,并且介绍了其数据储存结构与基本的检索方式。针对整个回环检测系统进行改进思路的阐述,首先是描述子提取方面,将原来的128维SIFT特征向量,改进为由角点检测、灰度值对比生成的二进制字符串描述子。并对回环检测在进行图像检索时的检索方式由原本的逐帧检测方式改进为将整个图像数据库人为分成高概率区间与低概率区间,对低概率区间进行隔帧检测缩短了检索时间。最后利用支持向量机算法将数据库中的图片进行分类训练,并建立预测函数,以改进整个回环检测流程的算法效率。最终,文章通过建立四个针对不同环境的独立数据集,并分别对原有算法与改进算法进行分组验证,首先验证了SIFT算法的图像特征检索能力,继而对比了SIFT描述子与改进描述子的性能,并将其应用到回环检测系统中。通过针对算法执行时间与算法的精度、召回率等评价指标进行分析,验证了改进后的FAST-BRIEF算子相较未改进的SIFT算子及SIFT-BRIEF算子其精度提升了37%、58%。运行时间方面使单张图片计算时间缩短10%左右,加入支持向量机对图像数据库进行训练和预测后,使得整个回环检测系统在保证其精度指标的前提下,系统召回率提升约44%。证明了该改进算法的优良性。更多还原

【Abstract】 Starting from professional automation area,robots become more and more common in modern industry and bring convenience to every aspect of daily life.The ability of self-localization and environmental maps building are the key points to create high intelligentized robot.SLAM(Simultaneous Localization and Mapping)is used to solve these problems with robotic vision system image processing.The main propose of this paper is to design a loopback detection system that corrects robots drift error caused by long running time and determines if robot has returned to a point in original path.Traditional loopback detection system extracts feature with SIFT algorithm,although it’s accurate,due to large volume of calculation,and RAM wasted by descriptors,this research improves the feature extraction algorithm and descriptors,and adjusts the query logic during loopback process,then tests all the improvements.First,this research analyses the architecture of Kinect camera used to collect images,describes the fundamental principles of TOF camera,introduces how Kinect calibrate,including the transformation of 4 coordinate systems and rectifies the camera used in experiments.Following that is the introduction of image feature extraction algorithm.Then steps of most frequently used FIFT algorithm is 1.Scale-space extrema detection 2.Key point localization 3.Orientation assignment 4.Key point descriptor generation.Their principles are explained in this part,as the fundamental of algorithm improvements afterwards.After that,the descriptors of photo features during loopback detection process are extracted.Feature matching and categorization are also introduced.Then,the principle of K-Means algorithm and bag of words are analysed.Data structure and basic search methodology are included.In terms of descriptor extraction,improving from 128 D SIFT eigenvector,the binary string descriptor generated from corner detection and contrast of grey range is used.And retrieving the image search mode when performing the loopback detection improved from the original frame by frame detection mode is divided into an entire image database artificially high probability and low probability interval range,low probability of detection range for spacer frame search time is shortened.Finally,the support vector machine algorithm is used to classify the pictures in the database,and the prediction function is established to improve the algorithm efficiency of the whole loopback detection process.Finally,the article establishes four independent data sets for different environments,and separately verifies the original algorithm and the improved algorithm.Firstly,the image feature retrieval ability of SIFT algorithm is verified,and then the SIFT descriptor,improved descriptor are compared.By analyzing the execution time and the accuracy-recall rate of the algorithm,It is verified that the accuracy of the improved FAST-BRIEF operator is improved by 37% and 58% compared with the unmodified SIFT operator and the SIFT-BRIEF operator.In terms of running time,the calculation time of a single picture is shortened by about 10%.After the support vector machine is trained and predicted by the image database,the system recall rate is improved by about 44% under the premise of ensuring the accuracy index of the entire loop detection system.It is verified that the performance of the algorithm is optimized and improved through improvement.更多还原