Object Detection-Based Visual SLAM Optimization Method for Dynamic Scene

In the realm of visual simultaneous localization and mapping (SLAM), the conventional assumption of a static environment presents notable challenges for achieving precision in dynamic settings. Moreover, the reliance on feature point-based techniques for generating sparse metric maps often proves insufficient for meeting complex higher level requirements. To tackle these challenges, we propose a real-time visual SLAM system that incorporates a parallel semantic thread for swift semantic scene analysis. By integrating a dynamic feature rejection module, our system enhances robustness in dynamic scenes by effectively leveraging both semantic and geometric information. A dense point cloud map is constructed using keyframes and their associated poses, employing techniques such as voxel grid filtering to mitigate map size. Experimental evaluations conducted on the public TUM RGB-D dataset and Bonn dataset validate the efficacy of our approach. The results showcase improved robustness and accuracy in dynamic environments, alongside the construction of dense maps for the static portions of real scenes, thereby facilitating downstream tasks.