An Localization Method Based Graph Attention Network For SLAM

With the rapid development of computer vision,the application field of visual SLAM has become more and more extensive.Due to the unstable change of the camera’s viewing angle during movement,the traditional visual SLAM has the problem of large positioning error and low detection loop accuracy.Visual odometer and loop detection are important modules of SLAM.Aiming at the problem of large positioning error and low detection loop accuracy,this paper proposes a visual odometer based on graph attention network,on this basis,this paper proposes a loop detection based on graph attention network and the two apply to SLAM.The experimental results proved the scientific validity of this method.The main work of this paper includes the following two aspects:1.Aiming at the problem that the change of camera angle of view is unstable,which leads to a large positioning error in the visual odometry module of SLAM,this paper proposes a visual odometry based on graph attention neural network.The traditional feature point matching algorithm in the traditional visual odometer is replaced with the feature point matching algorithm based on graph attention network,which is recombined with the traditional feature point extraction algorithm to achieve the goal of improving the estimated trajectory accuracy and pose accuracy,in the end,the camera can accurately locate.In this paper,the TUM data set is used for comparative experiments and the relative pose error and absolute trajectory error are selected as important evaluation indicators.Experiments show that when the camera angle of view is unstable,the visual odometer proposed in this paper is applied to SLAM,which can not only improve the accuacy of feature point matching,but also reduce the error of the estimated camera trajectory and camera pose,thereby improving positioning accuracy.Compared with the existing methods,the visual odometry proposed in this paper reduces the pose error by about 80%-90% by the maximum.2.Due to environmental factors such as illumination,angle of view,climate and other environmental factors,the camera angle of view is unstable,and the final SLAM has the problem of low accuracy of loop closure detection.This paper proposes a loop closure detection based on graph attention neural network.Analyzing the existing loop detection algorithms,this paper uses the graph attention network to improve the word structure of the bag of words model in loop detection,and directly adds the traditional feature point descriptor vector and feature point position vector to obtain the vector which is named a specific vector.And then the graph attention network is used to learn and update each specific vector to increase the specificity of each vector.You can use the final vector as a word.When searching for a word,you can rely on each words which are more distinctive and specific.The feature clusters are formed into words.All words form a dictionary.By calculating the similarity between the image and the image,the loop can be judged more accurately.In this paper,KITTI data set and self-made laboratory data set are used for comparative experiments.The calculation accuracy,recall rate and time performance are selected as evaluation indicators.Experiments have verified that in complex environments,the accuracy of loopback detection is up to 44%higher than that of existing algorithms.