Research On Location Retrieval Based On Fused 3D Laser And Visual Information

Location retrieval has been attracting wide arrange of domestic and abroad researchers' attention since 1940 s.In recent years,due to the powerful feature representation ability and learning ability of convolutional neural networks,people gradually realize that it is feasible and necessary to extract features using convolutional neural networks in location retrieval algorithms.At the Same time,3D laser information has been widely used in the field of intelligent driving,and the location retrieval algorithms based on 3D laser information have achieved great performance.However,location retrieval algorithms based on 3D laser information is easily disturbed by weather factors,and location retrieval algorithms based on visual information is susceptible to illumination variation.Therefore integrating 3D laser and visual information into location retrieval algorithm will help location retrieval algorithm overcome these environmental factors.This article main research work is as follows:(1)In order to solve the problem that the NetVLAD algorithm encodes all the local features of the input,so that the global vector of the output can not eliminate the interference of weak representative features,an improved NetVLAD,algorithm-MarginNetVLAD is proposed in this paper.Compared with the NetVLAD algorithm,the MarginNetVLAD algorithm contains redundant cluster points,which capture the local weak representative features through the improved margin lazy quadruple loss function proposed in this paper,so that the resulting global vector eliminates the influence of these weak representative features.(2)In view of the various problems existing in the loss function commonly used in the location retrieval algorithm,this paper proposes three new improved loss functions,and uses experiments to prove the advantages of these loss functions proposed in this paper compared with the commonly used loss function in the location retrieval algorithm.In addition,this paper applies the proposed MarginNetVLAD algorithm to visual location retrieval and 3D laser location retrieval algorithm,and experiments show that the performance of MarginNetVLAD-based location retrieval algorithm is better than that of NetVLAD-based location retrieval algorithm and other traditional location retrieval algorithms.(3)In view of the poor robustness of visual location retrieval algorithm to illumilation variations and 3D laser location retrieval algorithm to weather changes such as rainy days,this paper proposes four location retrieval algorithms based on fused 3D laser and visual information neural network according to the proposed MarginNetVLAD algorithm.For the weighted surface fusion algorithm,this paper improves the loss function and proposes a two-level retrieval scheme.In this paper,experiments show that the weighted surface fusion model is superior to other traditional location retrieval algorithms,and the recall rate of 1 candidate frames returned by retrieval algorithm is higher than that of NetVLAD algorithm 14.07%,higher than PointNetVLAD algorithm 18.01%.At the same time,experiments show that the weighted surface fusion model is more robust to the changes of environmental factors than other single information location retrieval algorithms.In this paper,two data sets are used to evaluate the algorithm proposed in this paper,and the generalization of the algorithm is verified.Finally,this paper analyzes the influence of various parameters on the performance of the weighted surface fusion model,and evaluates the algorithm complexity of the model proposed in this paper.