Research On Loop Closure Detection Of Mobile Robots Based On PCANet-LDA

With the wide application of mobile intelligent robots,simultaneous localization and mapping(SLAM)of mobile robots has become a hot topic for researchers.The loop closure detection part can bring strong constraints to the back-end pose map optimization,thus effectively reducing the cumulative error of pose estimation.It is an indispensable part of all SLAM systems.Principal Component Analysis Network(PCANet)can efficiently extract feature vectors of scene pictures,but it does not make use of the class discrimination of sample data.Linear Discriminant Analysis(LDA)can extract features using class labels in image data.In this paper,the advantages of PCANet and LDA are combined,and PCANet-LDA,a principal component analysis network with class differentiation,is proposed.Based on the theoretical basis of transfer learning,it is applied to visual SLAM loop closure detection.Firstly,the image is input into a two-layer PCA volume lamination layer to obtain PCA features.Then the feature vector is input to a nonlinear output layer for binarization and block histogram processing.Then the feature vector with class label information is input to LDA monitoring layer for monitoring projection,which makes full use of image data information.The experiment was carried out on the data sets New College and City Centre.Compared with other mainstream neural networks AlexNet,GoogLeNet and RandNet,the results show that the accuracy of PCANet-LDA is equal to that of GoogleNet under the condition that the recall rate is not more than 85%.However,the time for feature extraction is only 0.87% of that of GoogLeNet,and PCANet-LDA reduces the total time by 60.41% when calculating the similarity between feature matrices and the sum of the time for feature extraction.The experimental results show that the algorithm can effectively reduce the time cost of processing pictures while ensuring the accuracy.