为了在未知环境中进行导航，移动机器人需要构建环境地图并且同时定位自身在地图中的位置，像这样同时解决这两个问题的过程就称为同步定位与地图构建（Simultaneously Localization And Mapping，SLAM）。基于RGB-D传感器的SLAM也被称为RGB-D SLAM，是当前移动机器人研究领域的一个重要课题。RGB-D SLAM算法分为前端和后端两部分，前端算法对RGB-D图像进行特征检测与描述符提取，对提取到的描述符进行特征匹配，根据匹配结果估算运动变换并进行优化；后端算法根据前端算法的结果构建位姿图，然后进行闭环检测与位姿图优化，最终根据得到的最优相机位姿进行相机定位与三维地图环境重建。当前RGB-D SLAM算法的主要问题包括：首先，当前算法的效率低，不能满足实时要求；然后，当前算法的精度较低，误差较大，计算出的机器人位姿和运动轨迹通常会偏离真实值，并且随着时间增长偏离值会越来越大。针对现有算法存在的问题，本文对基于Kinect相机的RGB-D SLAM算法提出了以下改进方法：（1）在特征检测与描述符提取阶段使用基于ORB的特征检测和描述符提取方法，即使用ORB方法进行特征检测与描述符提取，并对深度信息不合法的特征点进行过滤。（2）在特征匹配阶段使用基于FLANN的增强特征匹配方法，即使用基于FLANN的KNN方法进行双向特征匹配，并且使用单应性矩阵变换对匹配结果进行优化。（3）在运动变换估计阶段使用改进RANSAC的运动变换估计方法，以得到更精确的inliers匹配点对。（4）在运动变换优化阶段使用基于GICP的运动变换优化方法，该方法基于高精度的GICP算法进行点云配准，使用inliers生成的点云进行GICP点云配准并对失败的情况做了退化处理，并且对RANSAC失败的情况做了退化处理，以提高点云配准的速度和精度。本文使用Freiburg提供的标准测试数据集对改进算法进行了测试与对比分析，并使用自己录制的实际环境测试数据和基于Dr Robot X80机器人的实际环境进行测试。实验结果表明本文所提出的RGB-D SLAM算法改进方法不仅能够满足实时性需求，而且能够极大地减小算法误差，提高算法精度，由此也证明了本文所提出的RGB-D SLAM算法改进方法的正确性。

To navigate in an unknown environment, a mobile robot needs to build a map of the environment and localize itself in the map at the same time. The process addressing this dual problem is called Simultaneously Localization And Mapping (SLAM). The SLAM based on RGB-D sensor is called RGB-D SLAM, and it is now an important issue in the research of mobile robotics.RGB-D SLAM algorithm is divided into a front-end and a back-end, whereas the front-end detects keypoints and extracts descriptors in RGB-D images, matches the descriptors to previous ones, then estimates the transformations and optimizes transformations, the back-end builds a pose graph, detects loop closures, optimize the pose graph, and then builds a 3D map of environment according to the optimized camera poses.Current RGB-D SLAM algorithm includes problems as below: first, it has such a low efficiency that it cannot meet the real-time requirement; second, it has a low accuracy and a large error, the resulting robot pose and trajectory will generally drift with respect to the real ones and the drift amount grows over time.Aiming at solving the problems mentioned above, this paper proposed the following improvement methods: (1) Use ORB based method in the feature detection and descriptor extractor stage, that is, use ORB algorithm to detect feature and extract descriptor, then filtered the features with invalid depth information away. (2) Use the enhanced feature matching method based on FLANN in the feature matching stage, that is, use duplex KNN feature matching based on FLANN library instead of just BruteForce method to match the descriptors and then using homography transformation to optimize the matching results. (3) Using an improved RANSAC method instead of original one to estimate transformations and get a more accurate inliers of matching points. (4) Using a transformation optimization method with automatic degradation based on GICP instead of original one to optimize the transformation, that is, use GICP algorithm instead of ICP to improve the accuracy of point cloud registration, treat the RANSAC failure with a degradation process, and use inliers matching points with a degradation process instead of the original 3D matching points to do the point cloud registration.This paper uses a benchmark and a corresponding dataset proposed by Freiburg and some RGB-D images sequences of different environments recorded by ourselves and the real environment based on Dr Robot X80 robot to evaluate the original algorithm and the improved algorithm, of which the result show that the improvement methods of the algorithm proposed in this paper can not only meet the real-time requirement, but also greatly reduce the errors and improve the accuracy. This proves the correctness of the improvement method proposed in this paper.