Research On Visual Place Recognition Based On Hierarchical Strategy

Visual place recognition is the core technology for localization and loop-closure detection in mobile robots and autonomous driving.However,changes caused by environmental conditions such as lighting,vegetation growth,and camera viewpoint lead to appearance variations and viewpoint variations at different degrees,which brings severe challenges to place recognition.To meet the major challenges,the paper proposes visual place recognition approaches based on hierarchical strategy,being robust against condition variations and viewpoint variations and meeting real-time performance.Specific contents are as follows:(1)In order to minimize the influences caused by image appearance change and viewpoint change in visual place recognition,a hierarchical visual place recognition based on semantic-aggregation is proposed inspired by the image representation based on aggregation.The process of visual place recognition based on single image is hierarchically reduced into two processes: candidate images extraction and place recognition based on semantic edges matching.Image representation with semantic significance is obtained by aggregation feature residuals between visual features and semantic features.Considering the efficiency and accuracy of matching,the semantic edges matching in different classes is introduced,to achieve an accurate place recognition in a small range.Experimental results show that visual place recognition based on hierarchical strategy has strong robustness against the seasonal change,day and night change,and viewpoint change,improving efficiency and accuracy.(2)Image sequences contain more comprehensive and rich information of one place compared to single image.Aiming at the problems of image appearance changes and viewpoint changes caused by environment condition and camera variations,an image sequence place recognition method based on hierarchical strategy is proposed,reducing sequence place recognition into two processes: image sequence candidates extraction based on first-to-end matching and place recognition based on sequence feature matching.The first-to-end matching is used to realize the rough localization of the query sequence on the reference database,and the scope of place recognition is narrowed.A sequence feature construction approach is proposed to excavate the overall information of the image sequence at a deeper level,to obtain the compact representation of the image sequence.Experimental results show that the approach based on hierarchical strategy can effectively reduce the computational pressure and has a strong ability to cope with severe appearance variations and viewpoint variations while maintaining good time performance.Meanwhile,the approach can still perform well even under extreme viewpoint changes.(3)In order to verify the effectiveness of the methods proposed,an actual environment database with severe appearance variations and viewpoint variations is collected by riding in the school.Experiments are carried out on the database,and the results show that the proposed methods are effective and have abilities to cope with the complex changes happened in the actual environment.