Research On Visual SLAM Algorithm In Indoor Dynamic Scene

Simultaneous Localization and Mapping(SLAM)based on vision is one of the core functions of autonomous mobile robots.At present,most visual SLAM systems are designed based on static environment,ignoring the influence of external dynamic objects on the system,resulting in large positioning errors and unable to operate normally.Therefore,based on ORB-SLAM2,a SLAM algorithm for indoor dynamic environment is proposed in this paper.The main work is as follows:(1)In order to solve the problem of too many weak information features due to uniform extraction of ORB features from quadtree,a method of ORB feature point extraction based on standard deviation was proposed.By dividing each layer of the image pyramid into equal blocks,the standard deviation is used to calculate the Hariss value of feature points,and the expected number of feature points are extracted from large to small according to the standard deviation calculation results.Experiments show that the improved method has higher feature matching accuracy than ORB algorithm and ORB-SLAM2 quadtree ORB algorithm.(2)Combining deep learning network YOLO-V3 with improved geometric constraints and introducing Bayes’ theorem,a dynamic feature elimination algorithm is proposed.YOLO-V3 and improved geometric constraints were used to filter dynamic feature points,and bayes’ theorem was used to judge the reliability of feature points.The lack of reliability of feature points caused by non-static environment is reduced and the positioning accuracy is improved.(3)The dynamic feature elimination algorithm is used to reduce the influence of dynamic objects,and the front end is used to screen out key frames to generate point cloud data,and the dense and octree point cloud map is constructed.The results show that the dynamic feature elimination algorithm can effectively solve the ghost problem in the process of ORB-SLAM2 diagram construction,and improve the robustness of the system in dynamic environment.