Research And Application Of Visual SLAM For Indoor Service Robots In Dynamic Scenarios

With the development of computer vision and artificial intelligence,intelligent robots based on visual SLAM technology are gradually serving our lives and attracted wide attention from academic and industrial circles.However,the existing methods have the following disadvantages:The traditional visual SLAM method makes static assumptions on the scene.Dynamic objects in the scene can interfere with position and attitude estimation and mapping.Existing methods use optical flow and semantics segmentation to remove dynamic object feature points in scene to eliminate interference,but fail to effectively integrate the two methods.In dynamic scenarios,the existing RRT* path planning algorithm can not quickly avoid moving obstacles and can not make full use of the previous planning information,so the planning efficiency needs to be further improved.In order to solve the above problems,this paper focuses on the dynamic feature point detection method based on fusion semantics and optical flow,and the planning efficiency of RRT in dynamic environment.The specific work is as follows:Firstly,a dynamic feature point detection method with tight coupling of semantics and flow is designed to solve the problem that visual SLAM is susceptible to dynamic object interference.Firstly,the influence of dynamic objects on position and attitude estimation of visual SLAM algorithm is analyzed,and the semantics segmentation network Mask is used.R-CNN divides potential dynamic objects in scene and proposes eight-point method for inter-frame motion estimation based on fused semantics;secondly,an optical flow dynamic feature point detection method based on dynamic threshold is proposed,which reduces the sensitivity of optical flow dynamic feature point detection to threshold value;secondly,a dynamic feature point determination strategy combining semantics and optical flow is proposed to make the semantics information and flow more sensitive to threshold value.Optical flow information complements each other and improves the accuracy of dynamic feature point detection.Secondly,an improved RRT* path planning algorithm is proposed to solve the problem of low efficiency of path planning in dynamic environment.First,according to the degree of collision risk between obstacles and robots,risk cost maps are introduced so that path planning can perceive the collision risk in advance.Second,based on RRT* path planning algorithm,the re-planning part is improved,useful paths in the initial planning are intercepted,and the previously searched paths are fully utilized to improve the efficiency of path planning.Thirdly,an experimental platform is built and experiments on pose estimation and path planning in dynamic environment are carried out in data sets and real scenes.The simulation results show that the positioning accuracy of visual SLAM system based on dynamic feature point detection in TUM data set is 14.5% higher than that of Dyna SLAM algorithm,and the efficiency of improved RRT* planning is 59.2% higher than that of traditional RRT*.Real-world experiment results show that the SLAM algorithm in this paper can maintain good performance in dynamic environment,improve the accuracy of position and attitude estimation of the robot greatly,and can effectively carry out path planning under the interference of moving objects.