PLPD-SLAM: Point-Line-Plane-Based RGB-D SLAM for Dynamic Environments

The majority of visual Simultaneous Localization and Mapping (SLAM) algorithms are built upon the assumption of static environmental conditions. However, this assumption limits the applicability of visual SLAM systems in real-world scenarios. Some methods use deep learning-based image segmentation to remove dynamic objects before tracking, which slows down tracking. Others relying on object detection end up with few point features left after removing those on dynamic objects, causing significant drift in the tracking trajectory. In this paper, we propose a dynamic SLAM method based on point-line-plane features. We calculate the information entropy to determine the distribution complexity of the pixels. If the information is sufficient, only point and line features are used for trajectory tracking, otherwise, plane features are added. We employ YOLOv5 for dynamic object detection, enabling robust tracking in dynamic scenarios by selecting reliable features. By performing object detection and tracking in parallel, we improve the real-time performance of the system. In sharp contrast to existing methods, the PLPD-SLAM can handle environments with dynamic objects in real-time and significantly reduce the long-term drift caused by dynamic objects. Finally, we evaluate our method using public benchmarks and our dynamic laboratory scenarios. The experimental results show that our method performs better compared to other state-of-the-art methods.