Research On SLAM Loop Detection Based On Convolutional Neural Network

Real time localization and mapping is the process of locating the target in the unknown domain and building the map of the surrounding environment,which is a hot research topic at present.Loop detection is an indispensable part of vision slam.Correct detection of loop can effectively reduce the cumulative pose drift of mobile robot in the process of positioning and mapping.At present,the traditional method of slam loop detection has good real-time performance,but the artificial feature is very sensitive to the change of illumination.To solve these problems,the method of deep learning came into being.In this paper,based on the method of deep learning,we continue to study the loop detection in slam system,and use convolution neural network to extract the required image features.In this paper,we propose a loop detection research based on deep learning.In order to solve the shortcomings of convolutional neural network structure and the lack of depth information in the input image,the innovation of this paper is to use two kinds of convolutional neural network to process single channel image,multi-channel image and depth image respectively on the input image for experimental comparison,and get a more robust slam loop detection algorithm.First of all,the convolution neural network is proposed to deal with it.This paper introduces the concept,method,application and implementation of loop detection in detail.The key point of this paper is to combine deep learning with loop detection.Secondly,the feature extraction of PCAnet convolution neural network is studied.Aiming at the problem that PCAnet convolutional neural network can not process RGB image at present,the image channel is processed.Through the results,we can judge whether the loopback occurs,so that the robot can make a decision and recognize whether it has reached a certain placeThen,for RGB-D In the process of image feature extraction,using vgg-19 convolution neural network,adding depth information to RGB image,that is,RGB-D image for convolution processing to extract features,two fusion measures of RGB image and depth image are proposed.By selecting the best fusion method,the image information is richer and the results are more accurate.Finally,in the above two experiments,the test and training before the experiment and the comparative experiment after the experiment are carried out to increase the experimental persuasion and eliminate the error caused by the contingency.The accuracy recall rate is used to evaluate the experimental performance,and the efficiency of the experimental results can be judged in time.The experimental results show that the improved algorithm has higher performance than the existing research methods,and enhances the efficiency of the algorithm.