语义线特征辅助的动态SLAM

动态环境干扰是视觉同时定位与地图构建（simultaneous localization and mapping，SLAM）领域内一个亟待解决的问题，场景中的运动对象会严重影响系统定位精度。结合语义信息和几何约束更强的线特征辅助基于传统ORB特征的SLAM系统来解决动态SLAM问题。首先采用深度学习领域的优秀成果SOLOv2作为场景分割网络，并赋予线特征语义信息；完成物体跟踪和静态区域初始化后，使用mask金字塔提取并分类特征点；再使用极线约束完成动态物体上点线特征的剔除；最后融合静态点线特征完成位姿的精确估计。在TUM动态数据集上的实验表明，提出的系统比ORB-SLAM3的位姿估计精度提高了72.20％，比DynaSLAM提高了20.42％，即使与近年来同领域内的优秀成果相比也有较好的精度表现。

Dynamic SLAM assisted by semantic line feature

Dynamic environment interference is an urgent problem in the field of visual simultaneous localization and mapping(SLAM). Moving objects in the scene will seriously affect the positioning accuracy of the system. This paper combined the line features with stronger semantic information and geometric constraints to assist the SLAM system based on traditional ORB features to solve the dynamic SLAM problem. Firstly, it used SOLOv2, an excellent achievement in the field of deep learning, as the scene segmentation network, and it could assign line feature semantic information. After object tracking and static region initialization, it used the mask pyramid to extract and classify feature points. Then, it used epipolar constraints to eliminate point and line features on dynamic objects. Finally, it combined static point and line features to complete accurate estimation of pose. The experiment on the dynamic data set of TUM shows that the pose estimation accuracy of the proposed system is 72.20% higher than that of ORB-SLAM3 and 20.42% higher than that of DynaSLAM. Even compared with the excellent achievements in the same field in recent years, it also has better accuracy performance.