Research On Loop Closure Detection Based On CNN

The loop closure detection algorithm based on bag of visual word(BoVW)can achieve satisfactory result in environments without significant lighting change and small-scale scene,however,it is difficult to achieve satisfactory results when lighting changes significantly.Recently,convolutional neural networks(CNN)has achieved great success in scene recognition and received much interest in applying CNN features to robotic fields.The accuracy of loop closure detection can be improved by learning the deep level unsupervised features through neural network.The parameter settings(network architectures,learning method,learning rate,etc.)and training time of large datasets of CNN applying in loop closure detection are analyzed and studied in this thesis.The main works can be described as follows:Firstly,the basic principles of bag of visual word and convolutional neural network are analyzed.And the loop closure detection algorithm based on bag of visual word is studied.Secondly,The pre-trained Places-CNN model is reconstructed,as a descriptor generator under Caffe.Thus,The CNN image feature descriptor acquisition method and the loop closure detection algorithm based on CNN are defined.Finally,The Caffe experimental environment is built under Linux system and simulation experiments are carried out.In the experiment,the loop closure detection algorithm based on BoVW is implemented by using VLFeat toolbox of MATLAB and Places-CNN was implemented by the open-source software Caffe.The pre-trained model of Places-standard scene dataset is retrained on two datasets of New College and Scene-15 datasets,and the trained model is used to extract image features and determine whether closed loop occurs by calculating the similarity between the current frame image and the previous frame image.The output characteristics of each unit layer of CNN is verified by the experiment.And the experimental results based on CNN closed-loop detection algorithm are compared with experimental results based on BoVW closed-loop detection algorithm,and two metrics is used to evaluate the performances:precision-recall curve and the mean average precision.The experimental results show that the CNN method is suitable for loop closure detection and the effect is better than method based on BoVW.