Research On Indoor Visual SLAM With Object Semantic Information In Dynamic Scene

Simultaneously Localization and Mapping as a key technology for robot autonomous movement.SLAM technology is used for robot to navigation and localization in industry,medical treatment and service field.The major dynamic target is slowly walk people,his carried object,and movable facilities.projection of dynamic object increase trajectory estimation error,result estimated trajectory of robot deviate from actual motion trajectory.Based-on semantic instance Visual SLAM System This thesis construct a global consistency semantic map which is based on static area and reduce trajectory error which is caused by dynamic object,Main works of this thesis are as follow:A kind of Visual Odometry used in dynamic environment is designed,The aim of this improved ORB-SLAM2 system is to solve the difficulty of inaccurate camera pose estimation,a dynamic object detection model is added in SLAM system.Dynamic object in the frame is recognized by Mask R-CNN instance segmentation network,dynamic object detection model is used to reduce camera pose estimation error caused by people and the object which is carried with people.different area pixel is described by binary encoding.The detected pixel was coded to 1 which belong to potential dynamic target.Rest of pixel was signed 0 which belong to static area.Then,The feature was extracted and matched from static area,iteratively calculate homogenous and fundamental matrix by RANSAC algorithm,geometrical relationship of frame is estimated;Finally,feature is searched and classified by geometrical relationship and binary encoding.After Visual Odometry estimate robot pose,robot construct environment map by estimated pose and motion trajectory,Different motion state feature is able to represent the different dynamic degree area.In order to represent real situation of environment,Optical flow vector is calculate to estimate dynamic degree of environment,K-means algorithm is used to cluster dynamic feature,then,point cloud is analyzed to extract its semantic information.Compared with others method,After feature encoding,semantic octomap retain feature of dynamic target to reflect dynamic target of dynamic degree in real environment.