视觉SLAM技术是机器人能够实现自主导航的基础之一，机器人利用视觉SLAM技术在不依赖先验信息的条件下，不仅能估计出自身位姿，而且能构建出当前环境的三维地图。目前很多成熟的视觉SLAM方案都基于静态环境完成定位和建图，然而实际场景中普遍存在的动态物体，会导致视觉SLAM定位精度下降以及地图产生重影，影响视觉SLAM技术在实际场景中的应用。针对室内动态环境下视觉SLAM定位精度低、建图噪声大的问题，提出一种基于ORB-SLAM2的改进算法。主要研究内容如下：

（1）基于LSD线与LT描述符的特征匹配算法。首先针对室内动态环境下ORB点特征有效数量少的问题，引入LSD线特征，丰富特征种类，对视觉里程计的特征提取进行辅助；然后针对LSD线常用的LBD描述子匹配鲁棒性差的缺点，引入基于Transformer的LT描述符构建LSD线的特征向量，利用欧式距离及最近邻准则进行匹配线对筛选。在Oxford、TUM RGB-D数据集以及真实场景中的实验结果表明，本文算法在相似纹理、低照度、光照变化、运动模糊、尺度变化的干扰场景中线特征匹配鲁棒性高，平均正确匹配数为57.8对，平均精度为87.38%。证明本文算法可以使视觉里程计获得数量更多、精度更高的特征匹配对。

（2）基于语义信息与点线特征融合的室内动态SLAM算法。针对ORB-SLAM2系统在动态物体干扰时定位精度差的缺点，首先引入轻量级SparseInst实例分割算法提取语义信息，然后结合极线几何约束进行实例级动态物体识别与滤除，最后利用静态区域点线特征融合进行位姿估计。在TUM RGB-D动态序列和自制真实环境数据集进行实验，结果表明本文算法能够有效滤除场景中的动态特征，相比于ORB-SLAM2，在w_xyz动态序列的绝对轨迹误差中，均方根误差降低了97.80%，标准差降低了97.92%。其余序列的绝对轨迹误差和相对位姿误差也有不同程度的降低。并通过消融实验、与DS-SLAM、RDS-SLAM算法的位姿估计结果进行对比、实时性对比，证明本文算法具有良好的位姿估计性能。

（3）室内动态场景下三维稠密地图构建。针对ORB-SLAM2用关键帧构建的稠密地图噪声多的问题，通过基于共视图的关键帧筛选、深度信息检测、动态像素消除、离群点滤波、体素下采样等去噪策略，生成低噪声的稠密点云地图，同时构建出八叉树地图。在TUM RGB-D动态序列和自制真实环境数据集进行实验，结果表明本文构建的稠密点云地图能够有效改善动态物体造成的重影问题。构建的八叉树地图冗余度低、占用存储空间小，适合应用于机器人导航且能满足大规模场景建图需求。

Visual SLAM technology is one of the foundations for robots to realize autonomous navigation. With visual SLAM technology, robots can not only estimate their own pose, but also build a three-dimensional map of the current environment without relying on prior information. At present, many mature visual SLAM schemes are based on static environment to complete localization and map construction. However, dynamic objects commonly exist in the actual scene, which will cause the localization accuracy of visual SLAM to decrease and the map to produce ghosting, affecting the application of visual SLAM technology in the actual scene. Aiming at the problems of low localization accuracy and high noise of visual SLAM in indoor dynamic environment, an improved algorithm based on ORB-SLAM2 is proposed. The main research contents are as follows:

(1) Feature matching algorithm based on LSD line and LT descriptor. Firstly, to solve the problem that the effective number of ORB point features is small in indoor dynamic environment, LSD line features are introduced to enrich the feature types and assist the feature extraction of visual odometer. Then, in view of the poor matching robustness of LBD descriptors commonly used in LSD lines, LT descriptors based on Transformer are introduced to construct LSD line feature vectors, and Euclidian distance and nearest neighbor criteria are used to filter matching line pairs. Experimental results in Oxford and TUM RGB-D data sets and real scenes show that the proposed algorithm has high robustness in matching midline features in interference scenes with similar texture, low illumination, illumination change, motion ambiguity and scale change. The average number of correct matching pairs is 57.8, and the average accuracy is 87.38%. It is proved that the proposed algorithm can make the visual odometer obtain more feature matching pairs with higher precision.

(2) Indoor dynamic SLAM algorithm based on semantic information and point-and-line feature fusion. Aiming at the disadvantage of poor positioning accuracy of ORB-SLAM2 system when dynamic objects interfered, lightweight SparseInst instance segmentation algorithm was first introduced to extract semantic information, then combined with polar geometry constraints to identify and filter dynamic objects at the instance level, and finally, point-line feature fusion in static area was used to estimate pose. Experiments on TUM RGB-D dynamic sequence and self-made real environment data set show that the proposed algorithm can effectively filter out dynamic features in the scene. Compared with ORB-SLAM2, the root-mean-square error and the standard deviation of the absolute trajectory error of w_xyz dynamic sequence are reduced by 97.80% and 97.92%. The absolute trajectory error and relative pose error of other sequences are also reduced to different degrees. The results of pose estimation were compared with those of DS-SLAM and RDS-SLAM, and the results were compared in real time to prove that the proposed algorithm has good pose estimation performance.

(3) Three-dimensional dense map construction in indoor dynamic scenes. In order to solve the problem that the dense map constructed by ORB-SLAM2 with key frames has too much noise, a low noise dense point cloud map is generated and an octree map is constructed through the noise reduction strategies based on common view key frame screening, depth information detection, dynamic pixel elimination, outliers filtering, voxel subsampling and so on. Experiments on TUM RGB-D dynamic sequence and self-made real environment data set show that the dense point cloud map constructed in this paper can effectively improve the ghost problem caused by dynamic objects. The constructed octree map has low redundancy and occupies little storage space, which is suitable for robot navigation and can meet the requirements of large-scale scene mapping.