DGFlow-SLAM: A Novel Dynamic Environment RGB-D SLAM without Prior Semantic Knowledge Based on Grid Segmentation of Scene Flow

Currently, using semantic segmentation networks to distinguish dynamic and static key points has become a mainstream designing method for semantic SLAM systems. However, the semantic SLAM systems must have prior semantic knowledge of relevant dynamic objects, and their processing speed is inversely proportional to the recognition accuracy. To simultaneously enhance the speed and accuracy for recognizing dynamic objects in different environments, a novel SLAM system without prior semantics called DGFlow-SLAM is proposed in this paper. A novel grid segmentation method is used in the system to segment the scene flow, and then an adaptive threshold method is used to roughly detect the dynamic objects. Based on this, a deep mean clustering segmentation method is applied to find potential dynamic targets. Finally, the results of grid segmentation and depth mean clustering segmentation are jointly used to find moving objects accurately, and all the feature points of the moving objects are removed on the premise of retaining the static part of the moving object. The experimental results show that on the dynamic sequence dataset of TUM RGB-D, compared with the DynaSLAM system with the highest accuracy for detecting moderate and violent motion and the DS-SLAM with the highest accuracy for detecting slight motion, DGflow-SLAM obtains similar accuracy results and improves the accuracy by 7.5%. In addition, DGflow-SLAM is 10 times and 1.27 times faster than DynaSLAM and DS-SLAM, respectively.