Place Recognition Based On CNNs And Semantic Graph

Place recogniton is an important task in computer vision and also a key component of loop closure in visual simultaneous localization and mapping(vSLAM).It is w idely used in fields such as robot navigation and autonomous driving.Place recognition under drastic viewpoint and appearance change is one of t.he most practical and difficult research fields.This paper proposes a scheme for place recognition under large viewpoint and appear-ance changes.Humans will confirm the similarity of the place semantic structure firstly and the similarity of texture distribution secondly when recognizing places.By emulating what humans do,we use the semantic spatial structure information and the image text.ure infor-mation distribution around the place,implementing robust place recognition under large viewpoint and appearance change.The main contributions of this paper are as follows:1)Since the good representation ability of neural network features in image descrip-tion,we develop a vie.wpoint-invariant and appearance-invariant SeqSLAM,which is based on neural network features and the hierarchical nearest neighbor algorithm.Compared to original SeqSLAM,the time complexity of this scheme decreases from O(N2)to O(Nlog N).2)Though scene appearance changes a lot in large viewpoint change,the semantic connection relationship will reserved.So a semantic weighted graph structure is proposed t.o describe the semantic structure of the place.3)Considering the matching of subgraph and subgraph is a NP-hard problem,we apply random walk method to graph embedding and implement a K nearest neighbors method to match the graph descriptor t.o find the right place.The time of the descriptor extraction and matching does not increase with the size of the graph,which can improve the efficiency of place recognit.ion.4)By emulating how humans recognize places,we propose a framework for place recog-nition based on semantic graph structure information and deep learning features.Based on the existed semantic matching results,we combine it with AD-SeqSLAM framework for more robust and accurate place matching.