| Positioning is becoming more and more important in people’s life.There are common positioning methods of GPS,WIFI,Radar,and Sensors,which are limited in use by signal coverage,and application cost.Computer vision-based spatial localization,on the other hand,has unique advantages in terms of environmental adaptability,simple equipment,and low cost,and thus has a wide range of application prospects.There are two types of vision-based localization,monocular and binocular,and monocular vision localization has lower equipment requirements and flexible deployment,and is more valuable for promotion.The method achieves localization by inverting the position and angle of camera shots based on matching the content of multiple images.Image matching is the key technology for monocular vision localization,and it needs to solve the problems of mis-matching,large computational effort,and low matching rate.The practical application scenarios such as small number of targets,small number of references and complex illumination make image matching more difficult and the above problems are more prominent.In this paper,for the image matching problem of monocular visual localization,we use the Attention mechanism to increase the attention to the locations with more valuable information in the data,so that the image matching algorithm can obtain more stable feature vectors and improve the matching accuracy.In location inversion,the GAN algorithm is used to expand the dataset of indoor application scenes to provide more reference points for the Pn P location inversion algorithm,and thus the amount,to improve the spatial localization accuracy and practicality.The algorithms proposed in this paper are evaluated and validated on both publicly available datasets and constructed indoor datasets.The main research work and results of this paper are as follows:1.Improvement of Key.Net image matching based on Attention mechanism.the Key.Net algorithm combines prior knowledge and shallow multi-scale CNN network to express human prior knowledge through neural network,but the importance of different channels is not highlighted in the multi-channel convolution process.In this paper,the Key.Net algorithm is combined with CBAM algorithm to improve the accuracy for feature vector finding.In addition,for the possible mismatching problem of image matching for visual localization,an iterative strategy is used to find more matching points in images with sparse feature points.2.Introduction of multi-scale image matching algorithm improvement.The Key.Net algorithm with Attention mechanism is better at fusing prior knowledge into the shallow CNN network,but it is not rich enough in multi-scale features due to the depth of the network.Inspired by the Gaussian pyramid in SIFT algorithm,this paper introduces multiscale information into Key.Net algorithm,i.e.,ATKey.Net.The model is trained on Image Net dataset and tested with HPatches benchmark,and the results show that the algorithm outperforms the baseline algorithm in terms of repeatability and matching performance.3.Location inversion method improvement.The data sample size has a great impact on the accuracy of location inversion.In this paper,the data used for position inversion is augmented based on the GAN algorithm and applied to the Pn P algorithm to improve the problem of low accuracy of 3D to 2D points due to the small amount of data.4.Visual localization application based on ATKey.Net image matching algorithm and Pn P algorithm.The proposed ATKey.Net algorithm and different image matching algorithms will be used to obtain feature vectors for the Aachen Day-Night dataset and the Pn P algorithm will be used for camera position inversion,respectively.Experiments show the superiority of the proposed algorithm over the baseline algorithm in terms of visual localization accuracy. |
