Research On Visual SLAM Algorithm For Mobile Robots In Dynamic Scene

Simultaneous Localization and Mapping(SLAM)technology helps realize autonomous operation of mobile robots in unknown environments.In recent years,the development of computer vision technology has promoted the rapid development of visual SLAM technology.However,although traditional feature-point visual SLAM has good accuracy in daily environment,this kind of method only considers static objects in the environment,and cannot deal with the influence of dynamic objects well,so its application scenario is limited to a certain extent.In view of this,this paper designs an efficient indoor dynamic SLAM algorithm based on the ORB-SLAM2 algorithm.The algorithm in this paper adopts the most simple and convenient monocular visual mode,which is applicable to most environments.The main contents of this paper are as follows:1)Based on the ORB-SLAM2,this paper designs a dynamic visual SLAM algorithm combining semantic information,optical flow method and geometric method.The algorithm uses improved Mask R-CNN instance to segment the network to obtain the contextual prior semantic information,and then remove the prior high dynamic objects.Then combined with optical flow method and geometric constraint algorithm,the polar distance relationship between matching points is used to detect and eliminate the real dynamic feature points in the environment.Experimental results show that the proposed algorithm can effectively improve the positioning accuracy of the system.2)In order to make the algorithm more suitable for indoor dynamic environment,this paper first slightly improves the network structure of Mask R-CNN to make it more suitable for indoor daily dynamic environment.Secondly,in order to solve the problem of reduced operating efficiency caused by the introduction of Mask R-CNN network to SLAM system,this paper adopts LK optical flow method to track feature points.By using optical flow method,FAST corner points can be quickly tracked without processing descriptors.This method has high real-time performance.Therefore,the optical flow method can be used in this paper to track the remaining feature points.This solves to a large extent the problem of system speed decrease caused by the introduction of instance split network.3)For dynamic features,this paper designs an efficient dynamic feature processing strategy.After the optical flow method is used to obtain the interframe information,the threshold is judged by the polar constraint,and the real set of dynamic feature points can be obtained.In this paper,the experimental results show that the dynamic object processing strategy in this paper is effective.4)Finally,the improved algorithm is tested comprehensively.The test results from TUM data set show that the algorithm in this paper improves the absolute pose error by about 9%compared with the ORB-SLAM2.And improved relative pose error by 39%.The experimental results show that the proposed algorithm can effectively reduce the impact of dynamic objects on visual SLAM algorithm in indoor daily living environment,especially in improving the accuracy of local pose of visual SLAM algorithm.