Research On Loop Closure Detection Technology Of Mobile Robot Based On Visual SLAM

Simultaneous Localization and Mapping is a significant technology for mobile robots to perform autonomous navigation in unknown environments.It has become one of the central issues which are studied by both domestic and foreign scholars.As an important part of the SLAM system,the technology of loop closure detection can eliminate the cumulative error of the pose estimation of the mobile robot after a prolonged movement,and ensure that the constructed map has global consistency.This paper studies the feature extraction algorithm and similarity measurement algorithm in the technology of loop closure detection.The research contents of this paper are as follows.(1)The system framework and mathematical model of visual SLAM are studied,and the working principles and usage scenarios of monocular cameras,binocular cameras and RGB-D cameras are analyzed;Feature extraction algorithms such as SIFT,SURF and ORB,dictionary generation algorithms and convolutional neural network models are studied.(2)The relationship between different dictionaries and the performance of the loop closure detection algorithm is studied,and the accuracy rate and recall rate of loop closure detection algorithm based on bag-of-words method are tested using datasets.The experimental results show that,when the train set used to generate the dictionary is more similar to the test set,the loop closure detection algorithm will achieve better performance,and that,to some degree,the scale of train set will also affect the performance of the algorithm.(3)The loop closure detection algorithm based on Inception?V3 network is proposed,the relationship between different similarity measurement algorithms and the performance of the loop closure detection algorithm is studied,and the accuracy rate and recall rate of the loop closure detection based on Inception?V3 network are tested using datasets.The experimental results show that,when using Inception?V3 network to perform loop closure detection task,PCA algorithm which is used for reduce the dimensionality of the descriptor vectors will improve the accuracy rate and recall rate of the loop closure detection algorithm,which is significantly better than the method without using PCA algorithm.The performance of the loop closure detection algorithm using Euclidean distance and cosine distance to measure the descriptor vectors are very close,and these two similarity measurement algorithms are both superior to the Manhattan distance.(4)The physical experiment used for testing the accuracy rate of loop closure detection algorithm based on bag-of-words method and Inception?V3 network is designed.The experimental results show that,compared with the loop closure detection algorithm based on bag-of-words method,the loop closure detection algorithm based on Inception?V3 network has a higher Accuracy.