Closed-loop Detection Of SLAM Based On Visual Dictionary Tree

With the continuous expansion of the robot exploration space,the Simultaneous Localization and Mapping(SLAM)system is especially important for improving the robot's perception ability.Closed loop detection is the key to building a robust SLAM system.At present,due to the robot's acquisition and description of visual information,there is a certain degree of error,which makes the SLAM system have low accuracy in image similarity judgment,and can not add correct closed-loop constraints,thus reducing the positioning accuracy in the unstructured environment.Based on this,this paper proposes a mismatch matching culling algorithm based on the improved Random Sample Consensus(RANSAC)algorithm to improve the correctness of the matching.Based on the similar scene detection,an improved image similarity algorithm based on dictionary tree is proposed.Thereby,the accuracy of the closed-loop detection of the system is improved and the positioning error is reduced.First,this paper studies the camera imaging model and calibration algorithm.The basic principle of small hole imaging and the transformation relationship between the four coordinate systems are emphasized.The camera is calibrated based on Zhang Zhengyou calibration method,and the point cloud information of color image and depth image registration is obtained.Secondly,the paper studies the mismatch matching culling algorithm based on improved RANSAC.Analyze the number of extractions,time spent,and distribution of common image features.Based on the feature matching mechanism,an improved RANSAC mismatching culling algorithm based on minimum threshold method is proposed.The minimum threshold method is used to eliminate the mismatch,and the RANSAC algorithm that reduces the number of iterations is used to further eliminate the mismatch.The number of correct matches between images is increased without increasing the time of eliminating mismatches.Thirdly,the paper studies the closed-loop detection of SLAM based on visual dictionary tree.The selection of key frames is carried out by using the motion between images,and the problems in the visual dictionary are small,and the ability to represent images is poor.First increase the number of visual dictionary tree layers to improve the visual word capacity of the image.Then use different levels of node words to judge the similarity between images,and then increase the accuracy of closed-loop detection.Finally,the closed-loop is verified by time consistency and polar geometry constraints,which suppresses the error closed loop and reduces the system positioning error.Finally,an experimental study of SLAM closed-loop detection based on visual dictionary tree is carried out.The closed-loop image with high similarity is selected by the dictionary tree,and the closed-rejection algorithm is used to eliminate the mismatch of the closed-loop image to ensure accurate matching constraints.By comparing the performance of the system and RGB-D SLAM in the dataset and real environment,the closed-loop detection accuracy and positioning accuracy of the system are verified.