Research On Loop Closure Detection Of Visual Simultaneous Localization And Mapping Based On Deep Learning

Simultaneous localization and mapping(SLAM)is mainly used to solve the problem of map construction and navigation of mobile robots in unknown environments,which is the basis for the autonomous movement of mobile robots.Because of the advantages of low cost,light weight,and abundant information acquisition,cameras have received great attention in recent years.Loop closure detection is a key step in visual SLAM,which can determine whether the current position has been visited by a mobile robot.Detecting the loop accurately can reduce the cumulative error effectively of the robot's pose estimation,and this is conducive to build a more accurate map and ensure the consistency of the generated map.At present,the method based on the model of bag-of-words has achieved great results,but the image features extracted by this method are sensitive to the change of illumination in the environment,and no loop can be detected when the change of lighting is significant.In recent years,deep learning has developed rapidly,many studies have shown that the features learned by neural networks are very robust to the change of lighting in the environment.However,neural network models in the field of loop closure detection often use supervised learning methods,and it require a large amount of labeled data to train the network,which has great difficulties in actual operation.What's more,the neural network extracts global features,which will lead to a low success rate of detecting loops when the image perspective changes greatly.In order to overcome these problems,in this paper,we propose the HSCAE model and HDLA model.The main works of this paper are as following:(1)We propose a spatial pyramid pooling module based on the convolutional autoencoding network(SPP-CAE),and design an unsupervised learning network HSCAE by combining the SPP-CAE module and the traditional methods of the histogram of oriented gradients(HOG).On the one hand,the local features of the image are learned through the HOG,and on the other hand,the global features of the image are automatically learned through the SPP-CAE.The image features learned by the two methods check and balance each other,so that the finally extracted image features can satisfy both the invariance to the change of illumination and viewpoint.Our method has the following advantages over existing methods.First,the model can directly train the network without the need for labeled image,which can effectively improve the effectiveness of training loop detection models.Second,the addition of the spatial pyramid structure enables neural networks to learn image features from multiple perspectives,which makes up for the deficiency that traditional convolutional neural networks can only extract global features of images.(2)We propose a loop closure detection model HDLA based on Hybrid neural network.Due to the features extracted from the middle layers in the neural network have good invariance to the change of viewpoint but do not perform well to the illumination changes,and the features extracted by the high layers of the neural network have good invariance to the change of illumination but do not perform well to the viewpoint changes.In addition,the features dimension extracted by the high level is too high,as well as the computational complexity.Therefore,we use HDLA to extract the high-level semantic features of key frames,and then use the local sensitive hashing method to reduce the dimensionality of the feature vector,thereby greatly improved the calculation speed of loop detection.The experimental results further prove that compared with the convolutional networks based loop closure detection methods,our algorithm improves the accuracy while ensuring the recall rate.