SD-SLAM: Visual SLAM Combining Semantic Segmentation and Dynamic Feature Point Detection

Most classical visual simultaneous localization and mapping (VSLAM) algorithms were researched on the basis of static scenarios. However, in the actual application environment, there are always some dynamic objects. Accordingly, a novel SLAM scheme based on semantic discrimination and improved dynamic feature point detection (SD-SLAM) is proposed which can achieve real-time localization and mapping via integrating the semantic thread and the corresponding semantic optimization thread. To be noted that the semantic thread and the localization thread are processed simultaneously and the localization thread does not need to await for semantic thread. In addition, the static feature point is taken usage for point cloud mapping and the non-prior dynamic feature point is adopted for dynamic feature detection condition. The extensive experiments for SD-SLAM on TUM benchmark dataset implies that the SD-SLAM achieves an improvement of 10.2% to 19.38% on absolute trajectory error and a decrease of 17.67
^%
on localization calculation time consumption for each frame, compared with SD-SLAM.