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

With the development of the field of mobile robots,the intelligent autonomy of mobile robots can be improved.It is a difficult research point for the mobile robot technology to detect the unknown environment and create real-time navigation.The mobile robot autonomously realizes its own positioning and establishes an environmental map based on its own position estimation and sensor data in an unknown environment.This process is called Simultaneously Localization And Mapping(SLAM).Vision sensors have the advantages of rich information,lightweight,and low cost.Combining SLAM with vision sensors has become a research hotspot for robot autonomous navigation.Visual Odometry can achieve the positioning of mobile robots,but it only considers the correlations on adjacent time frames,and the error caused by the motion of the mobile robot will inevitably accumulate to the next moment.,so that the cumulative error would occur in the entire SLAM.The results of long-term estimates would be unreliable,or we cannot build globally consistent trajectories and maps.If the loop closure detection is successful,the cumulative error can be significantly reduced and used as a basis for whether or not the map needs to be updated.It is of great significance to improve the robustness of the large-scale SLAM.The Bag-of-Word model is the current mainstream loop closure detection method.However,this method uses the characteristics of human design,in which human expertise and insights dominate the development process.This has great application limitations,and it takes a lot of time to extract characteristics.In the scene where the light changes significantly,the Bag-of-Word model method ignores many useful information in the environment,resulting in a low accuracy of loop closure detection.The purpose of deep learning technology is to learn representational data from raw data that can be used for classification.Loop closure detection is essentially a classification problem,which brings a new approach to the typical loop closure detection problem.This paper applies the latest vgg16-places365 convolutional neural network model applied to image classification and retrieval for the first time to visual slam loop closure detection.The paper uses this model as an image feature extractor,compares vgg16-places365 convolutional neural network model with other loop closure detection methods in the New College dataset.The result shows that the vgg model has achieved a good effect on the speed and accuracy of loop closure detection,so a new method has been explored for visual slam loop closure detection.It has certain innovation in the engineering field of mobile robot localization and navigation.The main work and results of this study are as follows:1.This paper describes the principle of each part of the visual SLAM framework,including sensor data,front-end visual odometry,back-end optimization,loop closure detection,and construction.The paper focused on the related principles of loop closure detection in detail.2.The mathematical model of the whole SLAM module is deduced in this paper and the state estimation is converted to the least squares method,and then the iterative method is used to solve the least squares problem,so that the whole SLAM problem is solved.3.In this paper,the feature point detection algorithm is studied,and the SIFT,SURF and ORB feature descriptor methods are analyzed and compared.The whole process of Bag-of-Word model is introduced in detail including the formation of the dictionary,calculation of similarity.At last we do the experiment to verify.In order to satisfy the strong real-time system characteristics of SLAM,the experiment uses ORB feature descriptors to extract features,which enhances the Rotation invariance and real-time performance of image feature matching;4.This paper first applies the vgg16-places365 convolutional neural network model to visual slam loop closure detection.This paper also visualizes the features extracted from each layer to make it more visualized.The framework of the network model of vgg16-places365 convolutional neural network is introduced in detail,then the training parameters of the model are introduced and then the loop closure detection method for experiment is given using the cosine similarity to calculate the two eigenvectors' similarity.Finally the paper gives the results and analysis of the experiment.This experiment is tested on the New College data set.Besides comparing with a number of traditional methods based on artificial design features such as BoVW,GIST,etc.,the paper also compares vgg16-places365 convolutional neural network model with the other several methods of deep learning model,and it was found that PR performance and feature extraction time performance are superior.