Research On Visual Place Recognition For Opposite Viewpoints Based On Multi-process Fusion

As an important method of loop detection in visual SLAM(Simultaneous Localization and Mapping),visual place recognition is a key technology in the field of autonomous mobile robots and autonomous driving.In visual place recognition for opposite viewpoints,there are fewer overlapping areas of the field of view,accompanied by occlusion problems caused by seasonal changes,illumination changes caused by the alternation of day and night,lateral displacement caused by lane changes,and common perceptual aliasing problems in place recognition,which increase the difficulty of place recognition.In order to adapt the system to the viewpoints changes in different environments,especially to maintain relatively high robustness in the case of opposite viewpoints,this paper proposes a visual place recognition method for opposite viewpoints based on multi-process fusion.The specific research content is as follows:Firstly,an efficient Coarse-to-Fine hierarchical place recognition structure is constructed for the visual images available in visual place recognition.In order to capture the information contained in the image,Refine Net is used for semantic segmentation,and a local descriptor Lo ST based on semantic aggregation is constructed based on the semantic information.The first T best matches are obtained through rough matching of descriptors using cosine distance.Then,keypoint mappings are obtained through the convolutional feature mapping of Refine Net,and filtering is performed based on semantics.Finally,keypoints are used for secondary matching of images based on spatial consistency,which is also known as fine matching,to obtain the final location identification results.Experiments conducted on the Oxford Robot Car dataset without lateral displacement show that the proposed algorithm can exhibit high robustness in scenarios with different disturbances.Secondly,aiming at the problem that the accuracy of the reverse viewing angle location recognition system is not high enough and its performance is poor in scenes with large changes in lighting,a visual place recognition algorithm based on multi-process fusion is proposed.Considering the high dimension and redundant information of the fusion descriptor,the concept of utility is proposed,and descriptor filtering is performed based on the cluster utility.A new descriptor L-Net VLAD is constructed to achieve more accurate rough matching.Considering the interference caused by dynamic objects in visual place recognition,a semantic-based keypoint secondary filter is designed to improve the accuracy of the fine matching stage.Experiments show that the proposed algorithm exhibits good coarse and fine matching performance on both the Oxford Robot Car dataset without lateral displacement and the Multi-Reverse Front-Rear dataset with lateral displacement.Finally,aiming at the problem of limited matching information for a single image,a multi-process fusion reverse perspective visual place recognition algorithm based on sequence and contrastive learning is designed.Firstly,a sequence-based matching algorithm is designed,and negative samples are mined based on the matching algorithm.At the same time,a triplet contrastive loss function for comparative learning is designed based on the matching algorithm,and descriptor optimization is performed through a trained contrastive learning model.At the same time,considering the training cost issue,the descriptor is dimensionally reduced based on principal component analysis before training,and then the reduced descriptor is used for contrastive learning training.Experiments show that the proposed algorithm can exhibit good place recognition performance in both coarse and fine matching on the Oxford Robot Car dataset without lateral displacement and Multi-Reverse Front-Rear dataset with lateral displacement.