Research On Closed Loop Detection Of SLAM Based Visual Dictionary Tree

Simultaneous Localization and Mapping(SLAM)is a hot topic in the field of mobile robot positioning and environmental cognition in the unknown environment.Visual based SLAM method has become a research trend at home and abroad.Visual sensor Kinect can collect image data and corresponding depth data,and can accurately perceive environmental information,and its price is low and is widely studied and used.At present,the classical vision SLAM algorithm using Kinect sensor is composed of image processing front-end and optimized backend.In the front-end visual odometer stage,the robot matches the current frame with adjacent frames to get the position and pose estimation of the current frame of the robot while the image is collected,thus the trajectory estimation of the robot in the environment is obtained.However,the errors produced by the robot in the last frame match will accumulate to the subsequent inter frame matching,and the result of the long-term estimation will be inaccurate,which leads to the inability to build a global consistent map.To solve this problem,a closed loop SLAM detection algorithm based on visual dictionary tree is proposed inthis paper.In order to solve the problem of cumulative error,this paper studies the front-end and back-end of RGB-D SLAM system respectively,and improves the scene modeling in closed loop detection,it obtains a kind of RGB-D SLAM algorithm with good accuracy and good robustness.First,the five modules of SLAM system are introduced in detail.The ORB algorithm and other feature extraction algorithms are compared and analyzed,and the ORB extraction algorithm is selected to meet the real-time requirement of SLAM;the adjacent images of the robot are matched with violent matching method,and the threshold distance is set to remove mismatch points.The robot trajectory is obtained through the RANSAC and ICP algorithm,and the cause of the cumulative error and its SLA are analyzed.The impact of M building maps.Secondly,the hardware structure,camera model and camera distortion of the RGB-D sensor Kinect2.0 for obtaining environmental information are studied,and the detailed process of correcting the distortion is described.The relationship between the RGB camera and the internal reference of the infrared camera and the relationship between the RGB camera and the infrared camera is calibrated.Finally,the color image and the depth image are fused.The results show that the calibrated Kinect sensor collects high degree of coincidence of pointcloud images.Then,the system framework of SLAM closed loop detection is elaborated,and the importance of scene modeling in closed loop detection is analyzed.The scene is modeled by Bo VW,and the visual feature list is formed by clustering algorithm in the scene.Because of the more visual words,in order to express the scene better,the visual dictionary can be better stored by the hierarchical structure,and it can be quickly indexed to similar historical frames.The classical Bo VW model uses K-Means to cluster,calculates the mean of the feature points in each cluster,and uses the calculation results as the cluster center.However,the K-Means algorithm is disturbed by the noise and other abnormal dimensions in the process of finding the cluster center,which will result in a large error between the actual center point and the center point which is computed.Therefore,the K-Mediods algorithm is proposed to cluster,and the K-Mediods algorithm randomly selects a feature point as a new center point from the class cluster generated by the iteration,avoiding the error caused by the characteristic points of the anomaly dimension in the average calculation.Finally,we use the open data set(Computer Vision Group)and data set evaluation tools to verify whether the closed loop structure and the algorithm before and after improvement.The results show that the closed loop detection can effectively reduce the cumulative errorgenerated by the visual odometer and the improved algorithm has a smaller absolute trajectory error,and the consistency of the map is better.In addition,building Turtlebot and Kinect sensor hardware platform for the actual indoor scene experiment,using the ICP algorithm to iterate the matching of the adjacent point cloud images,and continuously fuse the newly collected image information,splicing the image information into the generated local point cloud map,forming a three-dimensional map from the 3D point cloud scene to the whole indoor scene.The result verifies the improvement.The post RGB-D SLAM algorithm has high accuracy and robustness compared with the existing traditional algorithms.