Geometric Constraint-Based and Improved YOLOv5 Semantic SLAM for Dynamic Scenes

When using deep learning networks for dynamic feature rejection in SLAM systems, problems such as a priori static object motion leading to disturbed build quality and accuracy slow system runtime are prone occur. In this paper, based on the ORB-SLAM2 system, we propose method improved YOLOv5 combined with geometric constraint methods map building environments. First, paper uses ShuffleNetV2 lighten network, which increases network’s operation speed without reducing accuracy. At same time, pyramidal scene parsing network segmentation head is added part of achieve semantic extraction environment, so that has both target detection functions. order eliminate objects low features adopts constraints extract objects. By combining method, robustness interference targets construction eliminated. The test results TUM dataset show that, when constructing compared traditional algorithm, environment significantly improved. absolute trajectory error reduced by 97.7% ORB-SLAM2, relative position 59.7% ORB-SLAM2. Compared DynaSLAM scenes type, slightly improved, but maximum increase keyframe processing time 94.7%.