Research On Loop Closure Detection Based On Mobile Robot Visual SLAM

Simultaneous localization and mapping(SLAM)is one of the key technologies for mobile robots to achieve autonomous navigation in unknown environments.Because of low cost,low efficiency,light weight,the VSLAM research has become a research hotspot in the field of autonomous navigation for mobile robots.Among them,loop closure detection is one of the key in the SLAM problem,which is particularly important to eliminate the accumulated error of pose estimation and reduce mapping uncertainty.And mobile robots need to run in the environment for a long time,the collected image data is so large that the real-time requirements for loop closure detection are too high.Because moving objects may exist in the mobile robot's operating scene,the loop closure detection accuracy is often not higher.In view of the above problems,this paper conducts research on loop closure detection algorithms for mobile robots in complex indoor scenes.The main research contents include:1.The visual sensor mounted on the mobile robot is explained,the Kinect v2 depth camera's parameters,ranging principle and the corresponding imaging model is introduced in detail,the distortion and correction methods existing in the camera are explained,and detailed camera parameters are obtained through calibration experiments.At the same time,this paper introduces visual odometry in SLAM system in detail,including the process of extracting feature points of a given image using the ORB algorithm,using fast approximate nearest neighbor method to perform ORB feature matching on two images,and introducing random sampling consistency to filter the mismatches among them,the inter-frame camera pose is solved by the ICP algorithm.2.Aiming at the impact on the accuracy of pose estimation and loop closure detection when the mobile robot moves fast or there are moving objects in the robot's running scene,moving objects detection algorithm is introduced in this paper and the SLAM algorithm is improved.By using the initial camera pose estimation values to compute the motion vector of feature points at the current moment,the corresponding Gaussian mixture parameters can be calculated using the expectation maximization method.As the feature points will be included in the motion feature point set based on the result of the previous frame's motion point detection.However,by comparing the set thresholds,it is determined whether camera poses need to be solved again at the current moment based on the results of motion points detection.Experiments on given data sets show that the proposed algorithm can effectively detect dynamic points in scene images,and has better robustness.3.In order to improve the calculation efficiency of the algorithm and the calculation accuracy in dynamic scenes,firstly,the input image of the loop closure detection algorithm is pre-processed based on results of motion point detection,and images with a large dynamic feature points and high similarity with the previous image are eliminated.Secondly,the scene image is described by establishing a visual dictionary tree model,and the corresponding scene description vector can be given by using the image feature points whose motion pointe are eliminated.Then the similarity calculation method between two images in a single node is improved and by using the improved function,the occurrence of perceived ambiguity caused by similarity scenes can be effectively reduced and candidate loops can be obtained.Finally by using the temporal consistency and epipolar geometric constraint,false positive loops can be eliminated.The experimental results by using a variety of TUM public data sets and TurtleBot2 mobile robots show that the proposed algorithm in this paper can effectively detect the existence of loops in the scene.And it also proves that the proposed algorithm has higher accuracy in the pose estimation and a better mapping under different conditions.