Improved SLAM closed-loop detection algorithm based on DBoW2

In order to solve the problems of dbow2-based closed-loop detection algorithm, it is necessary to determine the appropriate level of dictionary tree and K value of clustering algorithm according to experience in advance. First, the new algorithm obtains the initial image feature clustering of training set by k-means ++ algorithm within the search range of cluster number (H, L). Secondly, by merging the form of similar classes, the cluster number is gradually reduced until the CRI function converges. Finally, the clustering value at this point is the best clustering value adopted by the training set to train the dictionary tree, and the clustering is repeated until the complete dictionary tree is generated. The feasibility of this algorithm is verified through experimental analysis, and the new method proposed can solve the problem of dictionary tree generation of training sets under different backgrounds, which plays a very important role in closed-loop detection based on visual image similarity detection, and has certain theoretical and practical reference value.