Feature Detection And Extraction Using GCNv2 And Image Mental Rotation Formobile Robot Visual SLAM

Feature detection and extraction for mobile robot visual SLAM is an important premise,which is directly related to robot pose estimation and environment map construction.Visual SLAM system has strict requirements for the selection of feature point algorithm under its real-time criterion: First,the detection and extraction of features can not occupy most computing resources;Secondly,the feature algorithm can deal with basic scene changes,such as image rotation,spatial scale changes and so on;Finally,feature tracking can still maintain a certain stability in bad or unexpected situations.In response to the needs of visual SLAM for feature algorithms,this paper designs a feature algorithm that combines the keypoint prediction of neural network features GCNv2 and mental rotation mechanism in neurobiology,called SIGCNv2 MR,and applies it to visual SLAM.In SIGCNv2 MR,GCNv2 is used to predict the keypoints in the image,that is,the coordinates of features in the image;The mental rotation mechanism is used to rotate the pixel neighborhood where the keypoints are located,and keep the BRIEF descriptor of the features unchanged when the image rotates.Furthermore,in order to ensure the scale invariance of features in dynamic environment and improve the speed of feature calculation,an image pyramid is constructed for the input image,and the keypoints are detected by FAST corners in the pyramid and their descriptors are calculated.We have verified SIGCNv2 MR for feature detection and extraction of visual SLAM system on mobile robot Xiaoqiang,which is called SIGCNv2MR-SLAM.The experimental results show that compared with the commonly used ORB feature algorithm,SIGCNv2 MR can extract uniformly distributed and dense feature points in the image,which improves the feature tracking ability of visual SLAM in the environment with few textures and pure rotation motion state,and image pyramid can overcome the scale distortion of 3-D point cloud hosting in the reconstruction of complex indoor constructions;In the process of robot moving,the mental rotation mechanism solves the difficulty that the descriptors of the same feature in two adjacent key image frames change after one frame rotates.Furthermore,we also tested SIGCNv2MR-SLAM on the international benchmark data sets,the experimental data show that SIGCNv2MR-SLAM has higher feature tracking robustness than ORB-SLAM2 and GCN-SLAM,and obtains more accurate motion trajectory in pose estimation,thus resulting in more accurate 3-D environment reconstruction.