Research On Loop Closure Detection For Visual SLAM Based On Deep Learning

In the unknown environment,Simultaneously Location and Mapping(SLAM)of mobile robots is one of the most important research topics in the field of robot navigation.Recently,with the development of deep learning technology and excellent performance in image processing,Visual SLAM based on deep learning technology has increasingly concerned.As a method to effectively reduce the error of environmental map,loop closure has become an important application in the field of deep learning.For the current study of Visual SLAM and deep learning algorithm,this paper analyzes the shortcomings of the traditional loop closure detection algorithm.For example,feature extraction is overly dependent on human cognition,the recognition of extreme image changes is not robust,recognition consumes a lot of computation and so on.Therefore,the deep learning algorithm is applied to loop closure detection in this paper,and proposes a new loop closure detection model based on convolutional autoencoder network.Then,in order to make the trajectory more accurate,the trajectory after loop closure detection is optimized.The specific research contents of this paper are as follows:(1)At first,we described the visual SLAM's overall flow system,tested and analyzed the traditional loop closure detection method.We find the limitations and shortcomings of the traditional algorithm.A new recognition model based on convolutional autoencoder network is proposed,which can be used for large-scale scene recognition.This model is characterized by two-channel end-to-end,feature embedding,good robustness and compactness.The feature extraction model of the model is built on the stack autoencoders,and then it is fused with the classic Net VLAD,which is called CAE-VLAD-Net(Convolutional Auto-Encoder – Vector of Locally Aggregated Descriptors – Network).We introduced the original Net VLAD model as the benchmark and effectively borrowed its high accuracy.In this paper,a wide range of experiments have been carried out under different challenging datasets to prove the recognition ability of the proposed model,and a comparison experiment has been carried out from the multidimension such as Precision-Recall curve?Area Under The Curve and feature extraction time.(2)Due to the continuity and density of sensor image acquisition,there are a large number of very similar continuous scenes,which will lead to a lot of error loops being detected.In order to improve the loop closure's accuracy,this paper proposes a feature extraction algorithm based on CAE-VLAD-Net model and applies it to the loop closure detection based on feature key frames.During the loop closure detection,according to the image feature distribution to determine the key frame position and then do processing.Instead of loop closure judgment for each frame,Euclidean distance of image features is determined based on key frame position and number of key frames.Finally,Determine the loop closure.(3)Loop closure detection is an effective way to reduce the error of constructing the map,but for the long-time movement of the robot and the extremely challenging scenes,there will still be errors in loop closure judgment.On the premise of judging the loop closure,we proposed a polymorphic incremental graph optimization method.this method regenerates the factor graph for the loop closure trajectory graph,the nodes in the factor graph are the set of several continuous poses in the closed loop graph.In which each node contains the posture of average state in a row.Finally,a new multi-state graph structure is constructed and the nodes are optimized under the graph structure.Polymorphic incremental graph optimization not only helps to generate more globally consistent maps,but also has a much lower optimization time than the approach that optimizes every pose node.