PIDNet-SLAM: A Multi-Resolution Semantic SLAM Algorithm for Dynamic Scenes

Simultaneous Localization and Mapping (SLAM) plays a vital role in the fields of Computer Vision and Robotics. Traditional SLAM frameworks often assume a static environment, which introduces significant errors during localization and mapping. To address this issue, researchers have incorporated various techniques to remove dynamic objects from the scene, but these methods fail to operate at real-time speeds. In this paper, we propose PIDNet-SLAM a semantic segmentation-based SLAM system optimized for dynamic environments. The system enhances the accuracy of ORB-SLAM3 by adding three parallel threads to the architecture: a fast semantic segmentation module using two branches of PIDNet for low and high resolutions separately, and a lightweight geometric module. We employed a novel training method to train the multi resolution PIDNet network, that utilizes the low-resolution branch to process non-keyframes and the high-resolution branch to process keyframes. This approach achieves high inference speeds while maintaining great accuracy. This multi-branch architecture significantly improves the performance of the semantic segmentation network, making it highly suitable for real-time applications. We evaluated our SLAM system on the TUM RGB-D dataset, comparing it with ORB-SLAM3, Dyna SLAM, and SOLO-SLAM. Our system achieved significant improvements, including enhanced localization accuracy in highly dynamic environments and an average processing speed of 47ms outperforming all other SLAM algorithms in handling dynamic scenes while operating in real-time.