Research On Semantic SLAM Algorithm Of Mobile Robot In Indoor Dynamic Scene

Simultaneous Localization and Mapping(SLAM)technology is becoming more and more mature,which brings more solutions for robots and UAVs.SLAM has entered the era of robot perception,and more development is needed to realize the effective understanding and perception of the environment for intelligent mobile robots.The point cloud map established by traditional SLAM cannot enable the robot to identify objects in the environment,because the traditional SLAM algorithm only contains geometric structure information,and its reusability is very limited.The dynamic target in indoor dynamic scene will affect the camera pose,so that the 3D map constructed will appear crossover,overlap and dynamic target.And combines SLAM and deep learning semantic SLAM can gift map semantic information,the robot can well understand the objects in the environment,people,as well as their relationship,and USES semantic information can eliminate dynamic target figure and the influence of the positioning accuracy,in order to satisfy the robot under the indoor scene can effectively achieve the task of autonomous navigation and decision-making to provide the necessary conditions.Abstract:Two semantic SLAM algorithms were designed to solve the localization and mapping problems in dynamic scenes,aiming at the interleaved and ghosting of 3D semantic maps constructed in dynamic scenes.Firstly,the visual odometer at the front end of visual SLAM is described,and its related mathematical basis and algorithm theory are expounded.In the current visual SLAM algorithm,the visual odometer will introduce dynamic feature points in the process of camera pose estimation in the dynamic scene,resulting in poor robustness of positioning accuracy.In this paper,we propose to use Mask R-CNN instance segmentation network semantic mask combined with multi-view geometry algorithm and Farneback optical flow algorithm to detect dynamic regions,and use this to reject dynamic feature points to optimize the camera pose of the visual SLAM algorithm in dynamic scenes.The experimental results demonstrate that the above method can ensure the rejection of most of the dynamic region features and improve the robustness and feasibility of visual SLAM in dynamic scenes.Then for building 3D semantic maps and solving ghosting caused by dynamic targets during map building,this paper uses semantic information from deep learning networks to build maps and solve the interference of dynamic targets on map building.The semantic information of Mask R-CNN is combined with a PCL point cloud library to construct a static 3D semantic octree map.Then the traditional visual SLAM combined with PSPNet semantic segmentation network is proposed to construct a semantic map,the camera bit pose is improved by Farneback optical flow algorithm as a way to eliminate the interlacing of dynamic scenes,the semantic information is fused with Bayesian semantics to construct a 3D semantic octree map by Octomap,and the predefined semantic colours of dynamic targets and Octomap's octree update method are used to Eliminate dynamic targets from the map and finally build a static 3D semantic map.The experiments are compared with a variety of mainstream visual SLAM algorithms for map building,and the semantic SLAM algorithm proposed in this paper has the ability to build 3D semantic maps in indoor dynamic scenes.