Loop Closure Detection And Optimization Of Visual SLAM Based On Deep Learning

In the environment,robots estimate their location based on data acquired from sensor and create environmental maps.This process is called Simultaneous Localization and Mapping(SLAM).It is the key to the implementation of autonomous mobile robots,and it has become a hot and difficult point in the field of robot research.Loop closure detection is a problem of scene recognition,and it is also the key link in SLAM system.It is very important to increase the pose constraints of the robot and reduce the cumulative error of the system.The traditional implementation method is based on the handcrafted features of the Bag-of-Words model detecting closed loop,its limitation is that it is affected by light and environment.Aiming at the above problems,we apply the deep learning technology in the aspects of scene recognition and feature extraction,and apply it to loop closure detection.A visual SLAM system based on deep learning to detect loop closure is proposed,which has strong theoretical and practical significance.For the visual SLAM front end,in the process of calculating the pose of the visual odometer,this thesis uses the pose calculated by RANSAC as the initial value,and then uses Bundle Adjustment to find an optimal value to obtain a more accurate pose estimation.Before the construction of the visual Bag-of-Words,the feature points and descriptors of the image are extracted by using the deep learning technique through the LIFT(Learned Invariant Feature Transform)network.The experimental results show that the LIFT features are more accurate and robust compared to other traditional handcrafted features.Considering the advantage of deep learning in scene recognition and feature extraction,this thesis proposes a closed-loop detection method based on the LIFT feature Bag-of-words model,and experiments are carried out on a public dataset.The experimental results show that the method proposed in this thesis is more discriminative in scene image recognition than SURF and ORB visual dictionaries.In the experiment of Precision and Recall,the loop closure detection results based on the LIFT feature show that the LIFT Bag-of-words model can achieve a higher recall rate when the precision rate is high.Finally,the feasibility of closed-loop detection of LIFT bag model was confirmed on the open datasets.In order to verify the theoretical research results in this thesis,this thesis combines the ORB SLAM system framework with the LIFT Bag-of-Words model to construct a visual SLAM system based on deep learning to detect loop closure.On the open dataset,a sparse point cloud,a dense point cloud and an Octomap map was built.On this basis,the performance of the improved SLAM system is evaluated.The experimental results show that the improved SLAM system has a good effect.Finally,this thesis summarizes the research work,and looks forward to the future research direction.