Research On 3D Map Construction Of Mine Environment Based On Machine Vision

As the depth of coal mining and the deterioration of mining environment accelerate,more and more accidents related to coal mining happen every year.In order to promote the efficiency of rescue and ensure the safety of rescue works,researchers try to make robots enter the roadway and participate in rescue operations.However,there is no GPS signal in the mine,which makes it difficult for the robots to walk autonomously.Autonomous navigation and localization is the key technology to realize intelligent detection and rescue,and the construction of 3D map is the basis of autonomous navigation and localization.Therefore,this paper introduces SLAM(simultaneous localization and mapping)into 3D mapping to construct a 3D map of the roadway in the mine,which can ensure underground emergency rescue smooth implementation.High-precision point cloud is the foundation of 3D map construction.However,the generation of point cloud map will produce a large errors due to the accuracy of the sensor and the external noise.The registration of point cloud will also produce a large cumulative error due to the accuracy of registration algorithm and a broadening view.A point cloud map occupies a large storage space because it has too much details,and the point cloud itself is discrete without topological structure,so it cannot support the rescue robot to walk and navigate independently.In view of the problems mentioned above,in order to obtain the point cloud information more easily,this paper adopts the RGB-D camera,which can collect image of information of depth and color at the same time,to collect information of the external environment.Aiming at the accuracy and timeliness of point cloud registration and the particularity of the roadway environment,we compared the applicability of SIFT,SURF,ORB in the roadway environment.After an experimental analysis,we finally selected the most robust ORB algorithm for feature detection.In the phase of eliminating feature point mismatches,the principle of the RANSAC algorithm was analyzed,and an improved RANSAC algorithm was proposed to eliminate mismatches.Instead of randomly selecting the matching feature points of the initial model,the quality matching features points were evaluated.The matching points with higher quality are used to generate the initial model,which improves the accuracy of the initial iteration data and effectively reduces the number of iterations.The experiment proves the superiority of the improved RANSAC algorithm,which reduces the time consuming and improves the accuracy of the mismatch rejection and initial registration.The accuracy of initial registration cannot meet our demand for high accuracy of the map,so we need to carry out a precise registration of the point cloud.However,it was found that some mismatches happened during the experiment by the ICP algorithm.Based on the analysis of the principle of the ICP algorithm,a new algorithm,namely the Two-Step Registration algorithm,witch takes the position of the point cloud into account and combines the advantages of 3D-NDT and ICP algorithms.The Two-Step Registration technology solves the problem that the ICP algorithm falls into the local optimal solution due to the large initial registration error caused by the variable roadway environment,and thus solves the problem of the ICP algorithm in the simulated mine environment experiment.In order to fix the cumulative error in the entire mapping process,we introduce local optimization based on key frames and Bag of Visual Words(BOVW)model to optimize the pose of the camera and to obtain a globally consistent map.In order to overcome the robots’ failure to walk autonomously in the generated 3D point cloud map caused by discrete point clouds,we converted the generated point cloud map into a 3D octree map to provide map services for the rescue robot.At the same time,the 3D octree map relieves our need for storage space because it occupies less.