Research On Visual SLAM Algorithm Of Mobile Robot In Dynamic Indoor Scene

Vision-based simultaneous localization and map construction(SLAM)algorithm is one of the core functions of autonomous mobile robots.However,most of the existing visual SLAM algorithms usually use external scenes as static assumptions for calculation convenience.The static system model constructed by them is sensitive to independently moving objects in the external environment,resulting in the SLAM system occur large errors in dynamic scenes,which cannot use in the real environment for the application.In view of this,based on the ORB-SLAM2 RGB-D framework,study a robust visual SLAM algorithm for indoor dynamic scenes,the core issues involved are ORB feature point uniform extraction,dynamic feature point detection and filtering,and static 3D map reconstruction.A multi-threaded SLAM framework is constructed based on ORB-SLAM2.A semantic detection thread was added to identify a priori dynamic targets in the scene.Added 3D map construction thread for static 3D point cloud and octree maps construction.In the tracking thread,the feature extraction module of the original system was improved,and a new dynamic feature filtering module was added to filter out the dynamic features in the scene.Aiming at the Quadtree-based ORB algorithm(Qtree_ORB)in ORB-SLAM2 has the problem of over uniform distribution of ORB features,an improved quadtree ORB features extraction algorithm is proposed.In the improved algorithm,the depth of quadtree splitting is limeited,features in each child node are retained by their Harris response values,which solved the problem of features over uniform distribution of standard algorithms;A grid-based ORB feature uniform distribution algorithm(Grid_ORB)is probosed.A grid division model is adopted to perform grid division on each pyramid layer,an adaptive threshold FAST features detection algorithm is used to extract feature on each grid of each pyramid layer,then a specific number of features in the grid are retained by their Harris respene values.Experiments show that features extracted by Grid_ORB algorithm have higher feature matching quantity and correct rate than Qtree_ORB algorithm.Aiming at the influence of dynamic features on the positioning accuracy of the SLAM algorithm,a motion consistency-based algorithm(MC)algorithm is proposed.A motion constraint geometric model is constructed according to the motion consistency between frames coarse filtering dynamic features.An improved RANSAC algorithm is used to calculate stable fundamental matrix,then truly dynamic features is removed by epipolar geometry;A semantic and geometric constraints algorithm(SGC),which combines deep learning network and multi-view geometry is proposed to filter out dynamic features in the scene.YOLO v3 is used to detect a priori dynamic targets on the current frame image,features on prior dynamic targets are filted firstly,and then the epipolar constraint method is used to filter out the ture dynamic features.The experimental results show that the MC algorithm and SGC algorithm can effectively filter out high dynamic features,and improve the positioning accuracy of the SLAM systemA static 3D point cloud map and octree map are constructed.Aiming at the problem of information redundancy of the map constructed by the key frames and dynamic targets appearing in the map,resulting in the map cannot be reused,a novel drawing key frame screening strategy is proposed.The dynamic information on the drawing key frames is filtered out,and then the depth data and the pose extimation matrix are used to calculate the spatial pose of the static pixels,completing the static three-dimensional point cloud map and octree map construction.Compared with other traditional methods,the information of the constructed map has less redundancy,and does not contain dynamic targets,which can be reused in the future.