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

With the proposal of "Made in China 2025",mobile robots are widely used in the field of intelligent manufacturing.As a key technology for mobile robot localization and navigation,visual SLAM has received a lot of attention.As the key link of visual SLAM,loop closure detection provides the core guarantee for the reliability of visual SLAM.The reliability of visual SLAM system without effective loop closure detection link is impossible to achieve.Traditional loop closure detection methods are mostly based on visual bag-of-words models,where the extracted image features are artificially set and rely on human experience,so the accuracy of traditional methods is not high and time-consuming.With the development of deep learning technology,convolutional neural networks have achieved good results in image processing,and these networks can be applied to loop closure detection.To meet the requirements of accuracy and real-time of loop closure detection,the following research is conducted in this paper.To address the problem of low accuracy of loop closure detection of traditional algorithms,this paper uses a modified lightweight convolutional neural network Ghost Net model to extract features by inputting the image to be measured into the Ghost Net after Image Net pre-training,and then uses the cosine distance to match the similarity of images to complete the loop closure detection.Experiments on the dataset demonstrate that the accuracy of the loop closure detection algorithm in this paper is the highest compared to traditional algorithms and other existing algorithms based on convolutional neural networks.Compared with the traditional method,the accuracy of the method in this paper was improved by 28% and 42.9% on the two data sets,respectively.To address the problem that the above loop closure detection algorithm proposed in this paper takes more time in the similarity matching,we further propose to combine this paper's method with the dimensionality reduction method.The self-encoder has good performance in dimensionality reduction,and the stacked sparse self-encoder has the best comprehensive performance after comparing with other self-encoder dimensionality reduction methods.In this paper,the high-dimensional feature vector to after Ghost Net feature extraction is then pre-trained with a stacked sparse self-encoder for feature dimensionality reduction,which reduces the dimensionality of the feature vector by 50% and reduces the time for subsequent similarity matching,so as to improve the real-time performance of the whole new loop closure detection method.On the basis of guaranteed accuracy,the algorithm time performance is improved by 19.1%and 16.9%,respectively,compared with that before the dimensionality reduction.In order to verify whether the method in this paper can meet the requirements of the visual SLAM system in a real scene,a visual SLAM system is built for the real scene verification of the method in this paper.After constructing 3D point cloud maps of the dataset images and laboratory scenes,it is verified that the method in this paper outperforms the traditional bag-of-words method and meets the requirements of the SLAM system for loop closure detection accuracy and real-time performance.