Research On Dynamic SLAM Based On Semantic Information And Multiview Geometry

Vision-based simultaneous localization and mapping technology(Visual SLAM)is considered to be an important basis for the intelligentization of mobile robots.This technology gives robots the ability to autonomously locate and build maps in an unknown environment.However,most of the current visual SLAM studies assume that the environment around robot is static.If there are large moving objects in the scene,the traditional method will be seriously interfered in the calculation process.The camera tracking error will increase and the map will have afterimage,which greatly limits the application of visual SLAM in real-world scenarios.Aiming at the urgent need of mobile robots for precise positioning and map construction in dynamic environment,this paper studies the camera positioning and mapping methods in dynamic scenes based on deep neural networks.At the same time,in order to solve the problem that the deeep neural network is inaccurate and the traditional method is greatly influenced by moving objects,We foucus on the feature extraction based on semantic information,the inter-frame pose estimation of point line features,and moving object detection based on multi-view geometry method,etc.what's more,We built a dynamic SLAM system based on RGBD sensor.Finally,the effectiveness of the algorithm is verified by the standard data set and real scene data.The specific methods proposed in this paper are as follows:Firstly,for the image feature extraction and matching problem in dynamic scene,this paper uses Mask-RCNN to semantically segment the color image to obtain the mask map,and then combines the geometric information of the depth image to refine the mask.In addition,the ORB feature points and LSD feature line segments are extracted from the color image and their descriptors are calculated respectively.Then the semantic information is used to classify and quickly match the extracted features.Secondly,aiming at detecting moving objects and improving the accuracy of pose estimation under dynamic scenes,this paper firstly calculates the pose of the current frame based on the feature points,and then uses the multi-view geometry method to calculate the space of the point cloud from the current frame.The consistency is detected,and the identified object is tracked by the median optical flow method.The tracking result is used to judge whether it is moving,and the above method is used to accurately segment the moving object.Then remove the pixels occupied by the moving object from the data of the current frame,and re-use the nonlinear optimization method to optimize the re-projection error of the dotted line feature to calculate a more accurate pose.In addition,the problem of ghosting of the map is solved by completely removing the pixels occupied by the moving objects during the construction process.Finally,We use the standard dataset and the data collected from the real scene to test the algorithm with the popular open source SLAM method as comparison,through several experiments to test the accurcy of localization and the quality of the map build in the dynamic scene.From the experimental result,we discover that the algorithm we proposed has big advantages in camera localization and dense map construction compared with the traditional method.