动态场景下基于自适应语义分割的RGB-D SLAM算法

An RGB-D SLAM Algorithm Based on Adaptive Semantic Segmentation in Dynamic Environment

目前较为成熟的视觉SLAM算法在应用于动态场景时,往往会因动态对象干扰而导致系统所估计的位姿误差急剧增大甚至算法失效。为解决上述问题,本文提出一种适用于室内动态场景的视觉SLAM算法,根据当前帧中特征点的运动等级信息自适应判断当前帧是否需要进行语义分割,进而实现语义信息的跨帧检测;根据语义分割网络提供的先验信息以及该对象在先前场景中的运动状态,为每个特征点分配运动等级,将其归类为静态点、可移静态点或动态点。选取合适的特征点进行位姿的初估计,再根据加权静态约束的结果对位姿进行二次优化。最后为验证本文算法的有效性,在TUM RGB-D动态场景数据集上进行实验,并与ORB-SLAM2算法及其他处理动态场景的SLAM算法进行对比,结果表明本文算法在大部分数据集上表现良好,相较改进前的ORB-SLAM算法,本文算法在室内动态场景中的定位精度可提升90.57%。

When the existing visual SLAM(simultaneous localization and mapping) algorithms are applied to dynamic environments, the pose error estimated by the system often increases sharply or even the algorithm fails due to the interference of dynamic objects. In order to solve the above problems, a visual SLAM system is proposed in this paper for indoor dynamic environments. By adaptively judging whether the current frame needs semantic segmentation according to the motion level information of the feature points in the current frame, the cross-frame detection of semantic information is realized. According to the prior information provided by the semantic segmentation network and the motion state of the object in the previous scene, each feature point is assigned a motion level and is classified as static point, movable static point or dynamic point.Some appropriate feature points(static points) are selected for initial pose estimation, and then secondary optimization of the pose is performed according to the results of weighted static constraints. In order to verify the effectiveness of the proposed algorithm, experiments are carried out on the TUM RGB-D dynamic scene dataset, and compared with ORB-SLAM2 and other SLAM algorithms for dynamic environments. The results show that the proposed algorithm performs well on most datasets, and the positioning accuracy in indoor dynamic environments can be improved by 90.57% compared with the ORBSLAM algorithm without the improvement.