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

SLAM technology is short for simultaneous localization and mapping.Another name of SLAM is CML,namely Concurrent Mapping and Localization.This technology is mainly to enable robots or UAVs to achieve autonomous positioning in an unknown environment and build a map of the environment.SLAM technology is undoubtedly a preferred choice in the process of robotic autonomy,because it can get abundant information according to sensors.And it has its greatly convenience for solving positioning and map building problems.This paper mainly studies the loop closure detection in visual SLAM.As we all know,loop closure detection is an important part of SLAM.It can not only make the constructed map continuous and uninterrupted,but also improve the accuracy of map construction.Compared with the traditional bag of word model,the loop closure detection method based on deep learning can improve the accuracy of loop closure detection.In this paper,we use the loop closure detection method of deep learning to extract the features of images in visual SLAM system by the trained deep learning model.And then calculate the similarity of images to get the loop closure detection results.Firstly,the depth camera is introduced.While acquiring the RGB image,the depth data of the corresponding image is obtained,which reduces the calculation amount in three-dimensional reconstruction of the surrounding environment.At the same time,the depth information in the environment is used to simplify the pose optimization algorithm,which is conducive to the real-time requirements of the system.Then the camera calibration theory is introduced and the camera is calibrated.Then the front-end design and back-end optimization of visual SLAM are analyzed and studied.After obtaining the position and pose information,the position and pose optimization is carried out.On this basis,the closed-loop detection of visual SLAM is studied.After the map is optimized by the closed-loop detection,the optimization results are returned to correct the overall map.Secondly,the traditional loop closure detection based on bag of word model is introduced in detail.And the process of constructing bag of word by robot after arriving at an unfamiliar scene is introduced.In this process,artificial features are extracted from pictures.These features are lexicalized to construct feature dictionary.Then,the image similarity calculating method based on TF-IDF in word bag model is introduced.The matching results are obtained by comparing the frequency of feature points in two pictures.The loop closure detection method based on the bag of word model is tested at finally.Thirdly,this paper introduces the loop closure detection based on deep learning model.And we mainly analyses the loop closure detection based on CNN network model and studies the loop closure detection based on auto-encoder network model.In order to reduce the dimension of features in CNN network and improve the efficiency of the algorithm,this paper optimizes the network by PCA dimension reduction algorithm.And then carries out the auto-encoder network.The auto-encoder network is introduced emphatically.By adding noise and sparse constraints to the optimization algorithm,the auto-encoder network has practical application significance.Loop closure detection based on deep learning model differs greatly from traditional bag of word model in the way of feature extraction.In the model based on deep learning,it is entirely self-extracting from the network,which improves the efficiency of image information utilization.Finally,the loop closure detection based on the bag of word model and the loop closure detection based on deep learning are compared in the data set and the actual scene respectively.By calculating the similarity matrix and the precision-recall curve,it shows that the method has certain practicability.It shows that the loop closure detection results in visual SLAM system can be improved on the premise of making full use of image information.At the same time,it is proved that the loop closure detection based on deep learning model is a reasonable and effective technical means.