Research On Loop Closing Detection Of Visual SLAM Based On Sequence Matching

As a key technology to realize the positioning and navigation of mobile robots in an unknown environment,simultaneous localization and mapping(SLAM)has been widely used in the fields of automatic driving,AR,VR and so on.As a key component of visual SLAM,loop closing detection give the backend a strong constraint for global correction to effectively reducing cumulative errors by identifying whether the robot returns to a previously visited position.The correct closed loop is critical to the globally consistent estimation for the SLAM system,but the wrong closed loop can have fatal results.Therefore,how to effectively improve the accuracy of loop closing detection while avoiding the wrong closed loop has become an important research direction.This paper analyzes the shortcomings of the traditional loop closing detection method,and conducts in-depth research on the scene image description and closed-loop decision model.Firstly,aiming at the shortcomings of traditional image feature extraction methods,and combined with the advantages of neural network in image representation,an image description method based on fusion CNN and VLAD features is proposed.Specifically,the pre-trained convolutional neural network is used to extract the abstract features of the scene image,and the feature map outputted by the middle layer of the neural network is expanded into a set of feature vectors to preserve the local spatial characteristics of the image,then the expanded feature vector is encoded using VLAD and reducing the encoded result by PCA.Finally,use the obtaining feature vector as an image description.For the closed-loop decision model,considering the time correlation of scene image sequences,a sequence matching decision model based on sparse similarity matrix is proposed.First,searching for the nearest neighbor K images of the current observation image in the map information through FLANN,and then consider the different speeds that the robot may have at the position of closed loop,use the greedy algorithm to greedily retrieve the possible closed loop positions,and construct a sparse similarity matrix.The possible closed loops are detected by the sparse similarity matrix.Finally,the validity of the above method was experimentally verified using a standard data set.Through the verification experiment of loop closing detection algorithm,it can be concluded that the proposed loop closing detection algorithm based on sequence matching can effectively improve the precision and recall rate of loop closing detection,and the recall rate is at least improved 30.08% when maintain 100% precision.The experiment of building map based on the proposed visual SLAM based on loop closing detection shows that the visual SLAM system based on sequence matching loop closing detection has better effect.Compared with other loop closing detection algorithms,the constructed point cloud map has no obvious missing,and the trajectory error is much smaller than other algorithms.