In most existing SLAM (Simultaneous localization and mapping) methods, it is always assumed that the scene is static. Lots of errors would occur when the camera enters a highly dynamic environment. In this paper, we present an efficient and robust visual SLAM system which associates dynamic feature points detection with semantic segmentation. We obtain the stable feature points by the proposed depth constraint. Combined with the semantic information provided by BlitzNet, every image in the sequence is divided into environment region and potential dynamic region. Then, using the fundamental matrix obtained from the environment region to construct epipolar line constraint, dynamic feature points in the potential dynamic region can be identified effectively. We estimate the motion of the camera using the stable static feature points obtained by the joint constraints. In the process of constructing environment map, moving objects are removed while static objects are retained in the map with their semantic information. The proposed system is evaluated both on TUM RGB-D dataset and in real scenes. The results demonstrate that the proposed system can obtain high-accuracy camera moving trajectory in dynamic environment, and eliminate the smear effects in the constructed semantic point cloud map effectively.
展开▼
