Research On Visual Slam Of Mobile Robot In Indoor Dynamic Environment

Robot navigation technology has advanced significantly in recent years as science and technology have progressed.As the core technology of robot navigation,Simultaneous Localization And Mapping(SLAM)is widely utilized in manufacturing,medical treatment,service,and other essential fields.At present,most SLAM systems are proposed under the assumption of a rich texture and static environment,while moving objects are inevitable in the real world.This thesis offers a semantic dynamic completion algorithm based on the Seg Net network as well as a novel RGB-D SLAM in a dynamic environment,targeting on resolving the issue that dynamic objects cause incorrect data association,resulting in a significant reduction of mapping accuracy.The specific research contents are as follows:Firstly,to solve the problem stated above,this research provides a semantic dynamic completion algorithm integrating the Seg Net network,which is used to segment the input image,extract the image’s semantic information,and process the potential dynamic object.The edge contour of the dynamic object is extracted via the edge detection algorithm so that as a judgment benchmark,the deep information will be used to complete certain semantically dynamic objects lost in the Seg Net network.Based on the completed image,the final image semantic information is acquired through adjacent semantically judgment.Secondly,the research proposes a visual SLAM algorithm combined with a dense and direct approach for the sake of higher accuracy of pose estimation in the dynamic context.Based on the semantic segmentation results acquired from the Seg Net network and that approach,the geometric and semantic errors are utilized as the minimum energy function with the Cauchy kernel function added to reduce the impact of the quadratic term of the anomaly in the error.With dynamic region elimination,spatial consistency of plane,and deep information screening,the dynamic coupling score of each pixel can be obtained,and different categories can be classified and processed.In the meanwhile,the initialization estimation of camera attitude can also be completed.The confidence of the camera post can be obtained by the spatial data association algorithm,which is then evaluated by the common visibility detection of the 3D map point set and the current frame,and whether the camera pose can be optimized by the feature point method.In this paper,the feature points of the semantic dynamic object area are removed,and the remaining ones are regarded as static feature points.To further reduce the wrong matching of feature points,the Geometric Epipolar constraint and RANSAC algorithm are used to remove some static matching points with the camera attitude taken as the initial pose.The Levenberg-Marquardt method is employed to solve the least square problem according to the re-projection error of static matching points,so as to acquire a more precise camera attitude.However,when the feature points come across low texture areas,the traditional fixed mode and keyframe tracking methods are not an effective approach for the camera’s accurate attitude,or even lead to a failure tracking of the visual odometer,resulting in the collapse of visual slam.Therefore,as the optical flow tracking should be added here,the LK optical flow method is employed to track the feature points extracted from the current image frame,reducing the influence of dynamic objects and improving the robustness of the system.Moreover,considering the camera pose and map points,the octree map is constructed in a static environment.In the process of construction,the two-dimensional latent semantic dynamic area is mapped to a three-dimensional space with dynamic objects further removed,so that a static map is built in the dynamic environment to eliminate the dynamic area in the plane and space.As a result,based on the simulation platform,this thesis compares the current algorithm with the traditional visual slam framework on the TUM dataset combined with dynamic processing,and the proposed algorithm can effectively improve the positioning and camera pose estimation accuracy of the RGB-D SLAM Algorithm in high dynamic and low texture regions.