Research On Place Recognition Of Visual SLAM For Mobile Robot

As an indispensable component of backend of visual SLAM for mobile robot,place recognition plays a key role both in eliminating drift of a map and robot relocation.But it has not been solved completely for a long time,especially in a scenario in which viewpoint and appearance simultaneously change.Recent advance works,which exploit ConvNet landmarks as representation of an image,have demonstrated promising performance under viewpoint and condition variations.This paper conducts work basing on detailed analyses of this classical approach to further elevate its robustness and searching speed.The dominant jobs of this work are summarized as follows:(1)This work comprehensively introduces the pipeline of a traditional ConvNet landmark-based visual place recognition and related algorithms,and analyses the characteristic representation method called ConvNet landmarks,along with the shortcomings of this traditional method.(2)A sequence search-based matching method is introduced here to replace traditional single frame way by using timestamp related prior.Sequence search-based approach applies the results of single frame matching and considers the similarity of two sequences as the scores of two corresponding central images.Besides,a ratio test is exploited to evaluating the discrimination of matched sequences,eventually loop closures are determined by the extent of discrimination.(3)A hashing-based landmark matching method is proposed to take place of the bidirectional nearest neighbor search that traditional method uses.Firstly,a hashing index is constructed off-line based on all extracted landmarks of all reference images.The theoretical time complexity of this indexing time is O(1).By this hashing index,globally optimal candidate matched landmarks can be fast obtained for a landmark of a query image.Then,a dominant set is established with all candidate images to reduce searching range of that query image.Additionally,a similarity calculation and evaluation method are set up for this special matched method.Extensive experiments were made to evaluate the efficiency of proposed methods of this paper.The two proposed methods in this paper perform excellently in datasets with viewpoint and appearance changes.The robustness of them exceeds the traditional method.Notably,hashing based approach improve the average searching speed by at least two orders of magnitude.