Deep Learning-Driven Real-Time SLAM Algorithm for Dynamic Scenarios

Most traditional Simultaneous Localization And Mapping (SLAM) algorithms are based on the theory that the scene is static, and this limitation restricts their application in realworld scenarios. In view of the problems such as poor positioning accuracy and insufficient robustness of traditional SLAM algorithms in dynamic environments, this paper proposes a realtime dynamic visual SLAM algorithm. Based on the ORB-SLAM3 framework, this algorithm innovatively introduces a semantic thread that operates in parallel with other threads, effectively solving the problem that the slow running of the semantic thread interferes with the tracking thread. The semantic information is updated and transmitted through the movement probability mechanism and stored in the map. Meanwhile, the data association algorithm is used to identify and exclude the dynamic points during the tracking process. Finally, through the test and evaluation using the publicly available TUM dataset, the results show that the proposed algorithm outperforms the existing algorithms in terms of robustness and real-time performance in dynamic environments.