【作者】 王玉全；

【导师】 赵国良；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 基于视觉传感器的移动机器人同时定位与地图创建是当前机器人技术研究中非常活跃的一个研究方向。全景视觉以其视角范围大、获取信息丰富的优点在移动机器人领域得到了越来越多的应用。通常对全景视觉图像信息的处理需要对其进行展开,算法复杂、实时性差,且基于视觉的同时定位与地图创建算法的时间复杂度过高,难以达到实时性。本文针对全景视觉的丰富信息难以准确和实时处理的问题,在不需要对图像进行展开的情况下,对基于全景视觉的移动机器人同时定位与地图创建方法进行了研究。首先,研究全景图像特征提取及匹配方法并对其进行改进。利用尺度不变特征提取SIFT提取全景图像特征。在全景图像的有效区域内,提出一种特征匹配的角度约束准则消除错误匹配;提出一种基于采样和Mean Shift算法的改进SIFT方法,在不需要对全景图像进行展开的情况下,利用控制采样点数量来控制特征点的数量,利用Mean Shift算法主动寻找尺度空间中的局部极值点。实验证明改进算法提高了特征提取与匹配的效率,提取出的特征点在图像序列中可以进行稳定而准确的匹配。其次,研究了基于全景视觉的移动机器人SLAM方法。给出四轮机器人的两轮差动简化运动模型,介绍了全景视觉成像原理与全景视觉传感器设计方法,并给出了本文使用的全景视觉成像系统参数。通过对全景成像过程的分析,建立了基于全景视觉的移动机器人SLAM系统感知模型,从全景图像的像素坐标中利用初始运动信息获得了特征点相对于移动机器人的三维位置信息。将运动模型和感知模型相结合,递推地得到了移动机器人同时定位与地图创建的结果。再次,研究了基于贝叶斯滤波的全景视觉移动机器人SLAM方法,将含有噪声的运动模型与含有噪声的全景视觉感知模型结合起来,通过迭代得到更准确的系统状态估计。实验证明贝叶斯滤波大大提高了系统状态估计的准确性,其中以FastSLAM的时间效率为最优。其应用的主要问题在于其数据关联的时间复杂度,通过对算法精度与时间复杂度进行综合分析,选取FastSLAM算法作为全景视觉移动机器的SLAM方法。最后,研究了基于全景视觉的移动机器人SLAM时间优化方法。利用改进SIFT方法减少SLAM过程中的特征点数量和特征提取匹配的时间,通过实验验证了该方法对SLAM时间优化的有效性和稳定性。根据全景视觉特征点的匹配次数与连续性,提出一种特征库的动态管理方法,提高了特征库中特征的利用率和SLAM算法数据关联的效率,为基于全景视觉的移动机器人SLAM提供了一种实时性较好的解决方案。更多还原

【Abstract】 The research on simultaneous localization and mapping (SLAM) of mobile robot with omni-directional vision is a very active subject now. Omni-directional vision sensor is more and more used in this filed because it offer a rich source of environment information in a wide angle of view. Usually, the omni-directional image is transformed to a normal visual image first, but it is a very complex process with a low efficiency. At the same time, the temporal complexity of vision-based SLAM is not real-time. So, the SLAM of mobile robot with omni-directional vision is researched in which the omni-directional image is not needed to be transformed.First of all, the scale invariant feature transform (SIFT) and modified algorithm for features extraction and matching in omni-directional are researched. An angle constraint is used to eliminate wrong matching in valid region of omni-directional images. A modified approach based sampling and mean shift algorithm is proposed to reduce the number of features generated by SIFT as well as their extraction and matching time. The features number is controlled by the number of sampling point, and mean shift algorithm is used to search local extrema points actively in scale space to improve the efficiency. It is demonstrated that the time of feature extraction and matching is reduced obviously by modified algorithm and the feature matching is steady and accurate.Secondly, the system of mobile robot SLAM with omni-directional vision sensor is researched. The motion model of four-wheel robot is simplified to two-wheel differential model. The principle and structure of the catadioptric hyperboloid omnidirectional vision sensor are researched, and the real parameters are calculated. The perceptual model is proposed by combining the omni-directional pixel coordinate and odometer data to get the three-dimensional coordinate in robot coordinate system. The SLAM result is achieved by combining the motion model and perceptual model iteratively. Thirdly, the of SLAM system based on Bayesian filter is researched. The motion model and perceptual model both noises included are combined to get an accurate system state estimation iteratively. It is demonstrated that the accuracy of system state estimation is improved by the uncertain information processing method, but the main problem is the temporal complexity of data association in SLAM. The FastSLAM algorithm is chose for omni-directional mobile robot SLAM because of its best temporal complexity.Finally, time optimization method of SLAM is researched. The modified SIFT is used to reduce the time of features extracting and matching in omni-directional image, and the number of features in SLAM is reduced too. It is demonstrated that the modified SIFT is steady and effective for SLAM time optimization. And a dynamic management method of feature database based on matching number and matching continuity is proposed. It is demonstrated that the utilization ratio of features and the efficiency of SLAM data association are both improved by the time optimization method, which is a good solution to keep real-time for the omni-directional mobile robot SLAM.更多还原