Visual Navigation Based On Local Feature Matching

Visual navigation is a hot research direction involving a wide range of fields.One of its applications is to use high-resolution images to restore scene information to meet people’s observation needs and enhance people’s sense of scene immersion.Existing visual navigation methods are mostly used in the field of robotics,which is a kind of real-time map construction and positioning.However,there are relatively few researches on visual navigation systems based on image sets and researches have been done mostly rely on larger-scale data sets and some prior tags such as GPS positioning.Therefore,there is still a lot of research prospects for visual navigation systems based on limited image sets.Considering the visual correlation between images can be described by the local feature matching,the scene can be reconstructed to realize visual navigation through the local feature matching.This paper mainly focuses on the research of visual navigation system based on local feature matching.The visual navigation system based on this matching algorithm can realize more effective multi-scene navigation.Multi-scene images can realize automatic scene classification,multi-view automatic classification,panoramic stitching of single or similar viewpoints,smooth browsing between panoramic images,and path planning.First of all,in the specific plan of the visual navigation system,this paper designs a global SFM algorithm based on the Infomap community division to estimate the camera position of multiple viewpoints in the scene.This method better reduces the influence of abnormal samples on the global SFM reconstruction effect.Next,this paper proposes a matching compensation maximum spanning tree algorithm,which is used to retain the loops with larger weights during the execution of the maximum spanning tree algorithm.In addition,the algorithm is used to automatically generate the stitching sequence of the input image for the NISwGSP algorithm,and experiments have proved that this method improves the stitching speed while ensuring the stitching effect.This paper also proposes a two-stage stitching scheme to deal with small parallax,which has achieved good stitching results.Finally,on the basis of the work of the previous chapters,this paper uses SIFT feature matching and uses the maximum spanning tree algorithm to achieve scene division.In the scene,the camera position information is obtained by realizing global SFM,combined with density clustering to achieve multi-view classification in a single scene,and Perform panoramic stitching on images obtained from a single viewpoint.The cluster centers obtained by camera clustering are connected using the minimum spanning tree algorithm,and the smooth browsing of panoramic images is achieved through the single-image relationship with the panorama.Design an important viewpoint screening method and combine with A~* algorithm to realize path planning.Experiments show that the method in this paper can perform better reconstruction and visual navigation planning for the scene described by a limited image set.