Research On Visual SLAM System Based On Optical Flow Method And Semantic Segmentation

Over the past 30 years,simultaneous localization and mapping(SLAM)technology has become increasingly mature,it has been widely used in robotics,automatic drive,smart home and other fields.However,for convenience,most visual SLAM algorithms assume that the environment is static,which limits the application of the algorithm in environments with dynamic objects.Due to the existence of dynamic objects in complex environment,the pose estimation accuracy of visual SLAM system will be affected,and it is difficult to meet the real-time requirements of the system.Therefore,based on DS-SLAM frame,this thesis proposes an algorithm that can adapt to complex environment.It mainly includes the following work:Firstly,after the system is initialized,we transfer every frame image to the tracking thread and semantic segmentation thread,the method of adaptive threshold is adopted to extract ORB feature points in tracking threads to adapt to different environment.Quadtree algorithm is used to improve the density of feature points.By limiting the iteration depth of quadtree algorithm,the extracted feature points can be allocated reasonablySecondly,the pyramid Lucas-Kanada sparse optical flow method is used to track the motion of corner points.At the same time,combining with the results of Segment semantic segmentation thread based on Caffe framework,it identifies highly dynamic objects and potential dynamic objects,and adds the detection process of small objects to judge whether the segmented objects are moving or not.Then,the dynamic feature points are filtered out by geometric constraints,and high-quality static feature points are reserved for feature matching.Finally,on the premise of image feature matching,the camera motion is estimated using Pn P(Perspective-n-point)in combination with 3d space points and projection coordinates,and then the least square optimization problem is constructed to adjust the estimated value and provide initial value for back-end optimization.A dense indoor semantic octree map is constructed by combining semantic information and mapping module.The TUM data set was used to evaluate the accuracy of this thesis algorithm and the existing DS-SLAM algorithm.Compared with DS-SLAM,the real-time performance of this thesis algorithm is improved by 9.02%.In the high dynamic data set,the camera pose error is reduced by 38.94%,and most of the errors are less than 0.05 m.In low dynamic data set,the error is less than 0.02 m.On static data sets,the performance of this thesis algorithm is as good as that of DS-SLAM algorithm.The experimental results show that this thesis algorithm not only maintains good pose effect in static environment,but also improves the performance in dynamic environment,and improves the positioning accuracy and real-time performance of the robot system in complex environment.