Research And Implementation Of Visual SLAM Loop Closure Detection Based On Siamese Network

Autonomous mobile robots are widely used in home services,smart business,in-telligent transportation,military reconnaissance and other fields.Simultaneous Local-ization And Mapping(SLAM)is a basic problem in the field of robotics.Loop closure detection is an important part of SLAM.Some of the loop closure detection methods are based on the Bag-of-Words model.Loop closure detection is performed using hand-craft features.However,the Bag-of-Words model based on hand-craft features has poor robustness,and it is difficult to extract effective features especially when encountering new scenes.In order to improve the adaptability of robot vision SLAM loop closure de-tection method,this thesis attempts to use learning-based features to realize loop closure detection.In the field of robot SLAM,to realize loop closure detection,we first need to ex-tract the features of images captured by the camera,and then determine whether the loop closure is detected by the similarity between the features.The traditional loop closure detection method is mainly based on the hand-craft features,such as SIFT,SURF,ORB,etc.These hand-craft features show limitations in practical applications,especailly when the features are not obvious or in unseen scenes.On the other hand,the learning-based features gradually show good adaptability in various applications.This thesis proposes a loop closure detection method based on supervised learning,referred to as Siamese-ResNet.It combines the advantages of Siamese network and ResNet.The main work is as follows:(1)Built a dataset for training and testing Siamese net-work(2)Designed network structure for efficiently extracting features(3)Establish an assessment mechanism to compare different methods.Through the results of the Siamese-ResNet and the Bag-of-words model on dif-ferent test sets,we can see that the loop closure detection method based on learning shows higher accuracy,better robustness and shorter time-consuming.The learning-based loop closure detection method is expected to make the visual SLAM applied to more scenes,thereby promoting the development and application of autonomous mobile robots.