DSP-SLAM: A Robust Semantic Visual SLAM for Dynamic Environments

Intelligent mobile robots have a wide range of applications in indoor scenarios and require the capability to perform simultaneous localization and mapping (SLAM) in dynamic environments. However, conventional SLAM systems face challenges such as non-robust feature extraction and prone-to-diverge localization in dynamic settings. To enable robots to truly understand the dynamic information within indoor scenes and accomplish advanced tasks, this paper proposes a robust semantic SLAM system that incorporates deep learning: DSP-SLAM. The system utilizes the neural network-based model for feature extraction and applies semantic segmentation algorithms to filter feature points on dynamic objects, thereby reducing the influence of moving targets and enhancing the robustness and accuracy of the SLAM system in dynamic scenes. Additionally, the system employs TensorRT technology for model acceleration to ensure real-time performance. We conducted comparative experiments with other mainstream SLAM systems on monocular, stereo, and RGB-D camera types using the TUM and KITTI datasets. The results demonstrate that DSP-SLAM significantly improves both absolute and relative trajectory accuracy, showing better robustness and precision in dynamic environments.