Dynamic Scene Vision SLAM Optimization Based on Semantic Segmentation Network

Aiming at the problem that the traditional real-time location and mapping (SLAM) algorithm does not have high accuracy in pose estimation in dynamic scenes, a visual SLAM algorithm based on semantic segmentation network in dynamic scenes was proposed. Firstly, the color map and depth map were collected by RGB-D camera, and the color map was input to the lightweight semantic segmentation network LR-ASPP to eliminate the a priori dynamic objects and obtain the semantic map at the same time. Then, the multi-view geometry algorithm was used to reject the non-prior dynamic objects, after separating the dynamic and static feature points, the optimized position and pose was obtained, and the pure static semantic octree map was constructed by combining the semantic map and depth map to improve the adaptability to dynamic scenes and directly used in the navigation system.The test results of the public dataset TUM show that the minimum absolute localization error of the algorithm in this paper is only 0.007 6 m. Compared with ORB-SLAM3, the localization accuracy in highly dynamic scenes is improved by more than 80%, and precise dynamic regions and accurate semantic maps can be obtained, which verifies that the algorithm has good localization accuracy and robustness in complex dynamic scenes.