Research On Closed-loop Detection Technology Of Robotic Visual SLAM

With the progress and development of autonomous driving,autonomous mobile robots and other technologies,how to accurately map the unknown environment and realize autonomous navigation has become an increasingly important research content,visual slam provides a good solution to this problem.In visual slam,the front-end is responsible for estimating the motion state of the camera,and the estimation value contains some error itself,after the robot moves for a long time,there will occur cumulative effect of error,which will lead to the final map being seriously deviated from the real environment,so it needs to be found and corrected in time.Closed-loop detection technology uses the similarity between images to judge whether the robot has reached a similar scene,and transmits the closed-loop information to the back end,so as to eliminate the cumulative error and correct the map.For the closed-loop detection technology in the visual slam system,the research in this thesis is as follows:(1)The working principle of five modules of visual slam was studied,including sensor data,visual odometry,closed-loop detection,the back-end optimization and map building;Focusing on closed-loop detection technology,a key frame selection method based on lk optical flow tracking is proposed to solve key frame selection problem.The method used lk optical flow tracking to track each feature point in the last key frame,counted them and calculated the proportion relationship between the number of tracked feature points in current frame and last keyframe.Based on this,it decided whether to insert the current image as a key frame into the queue.Experimental results showed that this method can effectively eliminate the influence of camera motion changes on the quality of keyframes.(2)The application of bow model in closed-loop detection was studied,aiming at image feature extraction,the thesis studies three manually designed feature extraction methods: sift,surf and orb.Through comparative analysis of experiments,orb was selected as the feature extraction algorithm in the bow model.To solve the problem of feature clustering in bow,the k-means algorithm was studied,and the feature descriptors were grouped into K classes to generate visual words.In order to improve the construction speed and search efficiency of dictionary,a k-branch tree was proposed to be used as the dictionary storage structure to store visual words.Aiming at the similarity measurement problem between feature vectors,this thesis studied three commonly used distance calculation functions.After comparative analysis,Manhattan distance was selected as the similarity calculation method.Finally,the closed-loop detection method based on bow model was experimented,and the effect of the algorithm was given by the PR curve.(3)To solve the problem of limited feature expression ability of traditional manual design,a combination of deep learning technology and closed-loop detection technology was proposed.The convolutional neural network technology was studied.Based on the analysis of Siamese network structure and related feature extraction technology,an improved Siamese cnn is proposed as a feature extraction network to generate feature description vectors,calculate similarity,and finally realized closed-loop detection.Experimental results showed that the proposed method is more accurate than bow method and Siamese network method.