室内机器人的同步定位与建图方法的研究

发布时间：2018-07-22 13:00

【摘要】：同步定位与建图(SLAM)技术作为机器人科学中的一个重要领域,在制造业,农业,医疗卫生业,服务业,国防工业等众多领域有着广泛的应用。其中室内机器人领域更是其中的一个热点。然而传统的室内机器人RGB-D SLAM由于在提取和匹配特征的时候计算开销大,所以算法的实时性并不理想。同时由于深度相机的结构特性,使得在边缘处的深度信息获取不完全。而大部分的特征都处于物体的边缘处,这种情况使得系统在计算机器人的位姿时,很难找到足够多的特征点,容易造成位姿的丢失。提出了一种结合显著性特征筛选的改进ORB算法。针对传统的RGB-D SLAM提取和匹配特征时计算量过大的问题,本文采用了ORB特征以减轻计算量,并通过盒状滤波器构建了尺度空间,使得ORB的尺度不变性得到了一定的提升。为了解决误匹配较多影响效率的问题,本文提出了一种结合视觉显著性特征的特征点筛选方法,本文采用基于颜色统计的高效空间显著图计算方法来计算每个关键点的显著值,同时通过特征点自身以及周围点的显著值来对特征点的匹配进行对比,将差距过大的点筛选除去,提高了匹配的精度和速度。提出了一种通过远近点分类和光束平差法计算RGB-D SLAM中的机器人位姿的方法。通过将特征点按照远近进行分类,对于深度可信的点直接获取深度;对于深度不可信的点采用多帧方式估算出深度信息,从而丰富地图中的关键点。通过光束平差法对空间点进行重映射构,用误差函数优化并求解出机器人的旋转和平移,同时结合改进ORB特征和显著性匹配,构建SLAM系统,使得位姿计算时对于边缘深度值的缺失有一定的鲁棒性。本文在TUM的SLAM公开数据集上进行了验证,试验结果表明,本文的方法可以有效地在室内环境下进行同步定位与建图。对于存在动态模糊的场景,本文的方法也可以取得很好的效果。
[Abstract]:Synchronous Positioning and Mapping (slam) technology, as an important field in robot science, has been widely used in many fields, such as manufacturing, agriculture, medical and health industry, service industry, national defense industry and so on. The field of indoor robot is one of the hot spots. However, the traditional indoor robot RGB-D slam is not ideal for its real-time performance due to its high computational cost in feature extraction and matching. At the same time, because of the structural characteristics of the depth camera, the depth information at the edge is incomplete. However, most of the features are located at the edge of the object, which makes it difficult for the system to find enough feature points in the calculation of the robot's position and posture, which can easily lead to the loss of the position and pose. An improved Orb algorithm combining salience feature filtering is proposed. In order to solve the problem that the traditional RGB-D slam has too much computation to extract and match features, this paper adopts Orb feature to reduce the computation load, and constructs the scale space through the box filter, which improves the scale invariance of Orb to a certain extent. In order to solve the problem that mismatch affects efficiency more, this paper proposes a feature point selection method combining visual saliency features. In this paper, the significant value of each key point is calculated by using an efficient spatial saliency map calculation method based on color statistics. At the same time, the matching of feature points is compared by the salient values of the feature points themselves and the surrounding points, and the selection of points with too large a gap is removed, which improves the accuracy and speed of the matching. In this paper, a method of calculating robot pose in RGB-D slam by far and near point classification and beam adjustment method is proposed. By classifying the feature points according to the distance and near, the depth can be obtained directly for the points with confidence in depth, and the depth information can be estimated by multi-frame method for the points that are not trusted in depth, thus enriching the key points in the map. The spatial points are remapped by the beam adjustment method, and the rotation and translation of the robot are optimized and solved by the error function, and the slam system is constructed by combining the improved Orb feature and salience matching. So that the position and pose calculation is robust to the absence of edge depth value. This paper is validated on the slam open data set of TUM. The experimental results show that the proposed method can be used to locate and map synchronously in indoor environment. For the scene with dynamic ambiguity, the method in this paper can also achieve good results.
【学位授予单位】：辽宁大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP242

【参考文献】