Research On Dense Mapping Of Multi-robot Based On Kinect

Simultaneous Localization and Mapping(SLAM)of intelligent robots is a popular direction for intelligent development today.After decades of development,single-robot visual SLAM has achieved many excellent results and has been applied in many aspects in reality,such as sweeping robot drones and autonomous driving.In the face of large and complex scenes,the actual tracking time of single-robot SLAM is relatively long,which directly leads to low efficiency of robot mapping.Therefore,multi-robot collaborative SLAM has become an inevitable choice.In this thesis,the main research is that a single robot builds a 3D point cloud local dense map based on ORB-SLAM2,and then transmits the data and integrates multiple sub-maps into a global map.The main research contents of this thesis are as follows:(1)First,the problem of constructing a local dense map by a single robot is studied.Through the Kinect camera to collect image data,tracking thread,local mapping thread,closed-loop detection thread,global optimization and dense map construction,a globally consistent 3D point cloud local sparse map and dense map are obtained.The sparse map is created by using the classic ORB-SLAM2 algorithm.For the dense map,a method is designed to obtain the key frames obtained in the construction of the sparse map,and then the dense map is constructed on the basis of these key frames.(2)Secondly,the problem of fusion of local maps of multiple robots into global maps is studied.This article focuses on the situation where two robots do not meet but walk through the same scene,and map fusion is performed by identifying the same scene.For the problem of how to identify the same area in the map fusion process,the word bag model is used to detect the similarity of the map.By importing the image dictionary tree and recording the vector descriptor of the image according to the kd tree of the dictionary,it is classified,stored and stored.Index number,and then you can perform scene recognition on the current image and the processed image.Then perform feature detection and matching on the obtained key frames.Due to the existence of noise interference,there are many mismatches in the matched point pairs.Random Sample Consensus(RANSAC)algorithm is one of the most important algorithms in the field of mismatch elimination,but due to its own uncertainty and low efficiency.The Progressive Sampling Consensus(PROSAC)algorithm uses the confidence ranking of feature points to filter out low-confidence points to reduce the sampling set to improve the accuracy and real-time performance of mismatch elimination.This article respectively describes and discusses the two in detail.Paired points after removing outliers are calculated by the Iterative Closest Point(ICP)algorithm to calculate the rigid body transformation parameters of the two local maps and perform Bundle Adjustment(BA)to optimize the parameters.So far,the two maps are transformed to the same coordinate,and the redundant information is removed by the elimination method,and the fused global map is obtained.(3)Finally,build an experimental platform and verify the algorithm with the data set and real indoor scenes.The choice of camera was discussed and compared.After confirming the camera,calibrate the Kinect camera.Data transmission between multiple robots is carried out in a centralized way under the robot operating system ROS(Robot Operating System).Analyze the experimental results of the single-robot SLAM generating 3D point cloud sparse map and dense map.The experimental results of the global map merged by the two mismatch elimination methods are analyzed.In this thesis,data sets and real indoor scene experiments are conducted on the research content to verify the effectiveness of the proposed method.The experimental results generally meet the expected goals.