Research On Dynamic SLAM Based On Deep Learning

Simultaneous Localization and Mapping(SLAM)is one of the key technologies in the fields of robot automation,autonomous driving,and augmented reality.It is the basic technology for intelligent devices equipped with sensors to perceive the external environment.When the main sensor of the device is the camera,we call it visual SLAM.However,most of the current visual SLAM research defaults that our equipment is in a static environment for map construction.If a large number of dynamic objects appear in the current scene for occlusion,the error of the mapping result will increase,causing the map to produce ghosting and routes.Offset and other issues,which greatly limit the application of visual SLAM in practical scenarios.In this thesis,aiming at the influence of dynamic objects on visual SLAM in the actual scene,by deleting the dynamic objects in the scene,the problems such as the decrease of system accuracy caused by dynamic objects are solved.In this thesis,the method based on deep learning is used to obtain the scene semantic information of the input image data,and the dynamic objects are recognized by combining the optical flow method and the bipolar geometric relationship.Mapping accuracy in the environment.The main research contents of this thesis include:(1)Investigate the impact of dynamic objects on the visual SLAM system,combine the deep neural network YOLACT to segment the original RGB image and its corresponding depth map at the pixel level,and construct a semantic information matrix corresponding to the size of the original image for the segmentation results..According to the result of instance segmentation,this thesis summarizes the segmentation categories into dynamic object category,potential dynamic object category and static background category.(2)In the feature extraction and matching stage,in order to improve the matching accuracy of dynamic object feature points,this thesis firstly extracts the semantic information of the original image,and then extracts the key feature points of the original image,and proposes a combination of optical flow and geometry.A semantic feature point matching algorithm for multi-coupling structures of information.The experimental results show that the method proposed in this thesis can effectively identify and process the feature points of dynamic objects in a complex dynamic environment.(3)In the mapping stage,dynamic objects are eliminated through semantic mask,and the loss of original image information caused by the elimination of dynamic objects leads to the problem of map data loss in the process of mapping.This thesis repairs the image based on the traditional machine vision method,restores the scene picture without dynamic objects,and finally establishes a semantic map.Experiments show that the method proposed in this thesis can effectively improve the localization and mapping accuracy of visual SLAM.In this thesis,by combining the deep neural network with the visual SLAM task,the deep neural network is used for instance segmentation of dynamic objects.Compared with the traditional dynamic visual SLAM algorithm,the method proposed in this thesis effectively improves the accuracy and robustness of the system.