Zhuo Pang,∗ Juwei Zhang,∗∗ Bo Liu,∗ and Benyi Ren∗∗
Simultaneous localisation and mapping (SLAM), SOLOv2, instance segmentation, dynamic scene, point cloud map
To improve the robustness and accuracy of the simultaneous localisation and mapping (SLAM) system in dynamic scenarios, this study proposes a parallel real-time system named STM-SLAM based on ORB-SLAM3. This system introduces a SOLOv2 semantic thread that runs in parallel with the tracking thread, which is used to obtain prior information of dynamic targets. By eliminating dynamic objects, the influence of dynamic feature points on pose estimation is excluded. Meanwhile, the threading building blocks programming library is utilised to process FAST corner points, and the real-time requirements of the SLAM system are met by accelerating the extraction of ORB feature points. In addition, combined with the semantic information obtained from the front end, a dense 3D point cloud map without dynamic objects is constructed through filtering and depth constraints. The system has been evaluated on the TUM Dynamic Benchmark Dataset. The absolute trajectory error has been improved by 95.30% compared with ORB-SLAM3, by 44.44% compared with DynaSLAM. In terms of time efficiency, it has been increased by 90.24% compared with DynaSLAM. Meanwhile, compared with other advanced dynamic SLAM systems, STM-SLAM has higher positioning accuracy and greater system robustness.
Important Links:
Go Back
