Research And Application Of Edge-based SLAM System In Dynamic Environment

With the development of unmanned vehicles,service robots,AR/VR and other fields,Simultaneous Localization and Mapping(SLAM)based on visual sensors has become a hot topic in the field of artificial intelligence and robotics.However,in dynamic environment,the classic SLAM algorithm based on static assumption will have some problems,such as poor performance of motion estimation and low quality of environmental map.Therefore,how to improve the visual SLAM algorithm to achieve high-precision pose estimation in dynamic environment,and at the same time build a map that can guide mobile robots to realize path planning or navigation function is a subject with practical value.Aiming at the problem of poor robustness of edge-based visual SLAM algorithm in dynamic environment,an improved algorithm is proposed from three aspects: frontend visual odometer,back-end loop detection and relocation and mapping module,and the improved SLAM system is applied to the path planning of mobile robots.The main work of this dissertation includes:(1)A visual mileage calculation method based on dynamic environment is proposed.Firstly,the position of dynamic objects is obtained through semantic network,and the edge point clouds are clustered,and the clustering clusters that may be dynamic objects are marked by semantic prior.Then,the re-projection error of edge pixels in each cluster is calculated,and the dynamic region in the image is further judged according to the projection error and depth change.Make full use of the image classification results of semantic network,construct semantic projection error,and weight it with edge projection error to construct optimization function.The experimental results show that the visual mileage calculation method for dynamic environment designed in this paper can resist the interference of dynamic objects,and has good pose estimation accuracy while successfully removing dynamic objects from the image.(2)A loop detection method using edge point cloud for relocation is proposed.Considering the robustness of image edge to illumination change and motion blur,after downsampling the edge point cloud,the description vector of local edge point cloud is extracted by convolutional neural network,and a loop detection and relocation method based on point cloud descriptor is designed.The experimental results show that more trajectory loops can be detected by using three-dimensional local description vectors,and the positioning accuracy and long-term stability of the system can be improved.Based on RGB-D camera,point clouds are spliced and filtered to construct dense point cloud map and octree map.(3)Integrating the improved algorithms in the first two chapters,the SLAM system based on edge in dynamic environment is realized,and the dense point cloud map and octree map are established by using this SLAM system in real dynamic environment,and the octree map is projected to a two-dimensional plane to construct a grid map,so as to guide the robot to plan its path.The experimental results show that the SLAM system designed in this paper has a good performance in the actual dynamic environment,and the established environment map can guide the path planning algorithm and realize the navigation of the mobile robot,which proves the practicability of the algorithm in this paper.