Dynamic object removal by fusing deep learning and multiview geometry

SLAM technology involves using sensors mounted on vehicles to capture external environment data, to estimate their position in the environment and build a map model around them. Most of the existing SLAM solutions are developed under the assumption that the environment is static, but this assumption faces many challenges in dynamic environments. To reduce the negative impact of dynamic objects on positioning accuracy, this study proposes a dynamic object elimination technology combining depth learning and multiview geometry. This method first uses a semantic segmentation network to identify and remove the known dynamic elements in the scene, and then multi-view geometry is used to further exclude the actual dynamic objects. On this basis, a local map strategy is added to the ORB-SLAM2 feature matching algorithm to screen more representative points. For unable to match the feature points, this study adopted a method of feature point spread and used a high degree of confidence in the area of the weighted calculation of the extension point, thereby gaining the status information of the feature points.