Visual SLAM Research On Mobile Robots In Indoor Dynamic Environments

Intelligent robot technology has broad application prospects,and SLAM(Simultaneous Localization And Mapping)technology is the core of research in the field of intelligent robots.With the development of computer vision and sensor technology,visual SLAM based on RGBD camera has attracted extensive attention.At present,the research of visual SLAM mainly focuses on static scenes,and the research in the dynamic environments has not yet formed a unified scheme.Most dynamic visual SLAM have lower operating efficiency.Aiming at these problems,this thesis studies the visual SLAM in the dynamic environments with fixed dynamic object categories.The main contents are as follows:The contrast experiment of the mainstream feature extraction algorithm was carried out.The quadtree structure was used to improve the ORB feature extraction algorithm,and the uniformly dispersed feature points were obtained,which reduced the local feature point concentration.A visual front-end tracking algorithm based on deep neural network optimization was proposed.The YOLO v2 network was trained through migration learning.The network was used for dynamic object detection and the feature points in the dynamic object range were well removed,which reduced the influence of dynamic objects for matching and camera pose estimation.The keyframe-based closed-loop detection and test experiments were carried out.The common datasets were used for comparison,and a good closed trajectory was obtained.The visual backend optimization was carried out by using local BA optimization and global pose optimization,and more accurate positioning results was obtained.ORB-SLAM2 and DynaSLAM were compared under the public datasets about dynamic scene,the results show that the proposed algorithm had reliable positioning accuracy and had a higher improvement in operating rate than DynaSLAM.And,a mobile robot platform was built to experiment in real indoor dynamic scenes.The real-time positioning and composition tasks were completed.Experiments showed the feasibility of the SLAM algorithm running in real time in dynamic environments.