Place Recognition And Loop Closure Detection Of Mobile Robots Using Lidars

With the development of unmanned platform technology,its perception and navigation problems have begun to be seen as the key to full automation without human interference,among which place recognition and loop closure detection algorithm is the fundamental module of the navigation system.Especially in the field of simultaneous localization and mapping(SLAM),in order to eliminate the drift of the front-end odometry and construct a map with spatial consistency,it is necessary to integrate the place recognition and loop closure detection module.The current mainstream solution for place recognition is based on 2D vision,but the camera itself is greatly affected by the changes in lighting conditions and has a limited field of view,which limits its use in complex environments such as long-term,multi-field and crossseason situations.In contrast,the 3D Li DAR sensor is not affected by light changes and can accurately obtain distance information with a wide range of measurements,which makes it a great substitute for the camera to complete place recognition tasks in complex outdoor environments.At present,the 3D vision based on Li DARs is still premature,the processing of3 D data,feature extraction and semantic analysis are still open issues.In this context,an indepth study of the place recognition and loop closure detection of outdoor mobile robots based on Li DAR is carried out in this paper.The specific contents are as follows:For the problem of environmental understanding of outdoor mobile robots based on 3D Li DAR,a method for semantic classification of point cloud objects based on deep learning is proposed in this paper,which extracts objects in the environment and classifies them semantically.This method uses the clustering segmentation technique to segment scene-level point clouds into object-level point clouds to reduce the difficulty of learning.A deep neural network that can directly process 3D points is constructed based on one-dimensional convolution and symmetric function.The network does not need conventional operations such as projection and grid division,so it can make full use of the structural information of the point cloud.The segmented objects are classified into four categories according to their semantic characteristics: car,person,cyclist and other.The semantic classification network is trained to realize the semantic understanding of 3D point clouds.Experiments on KITTI datasets prove that the proposed method has good accuracy and real-time performance.To solve the difficult problem of place recognition based on 3D point clouds in large-scale,highly dynamic and complex outdoor scenes,the sparse semantic feature map(SSFM)is proposed and used in this paper for place representation and place matching with coarse to fine registration.Each point in the sparse semantic feature map includes the geometric center of the point cloud object with semantic label and high-dimensional feature.The class information and object feature are acquired by a deep neural network trained based on the multi-task learning strategy.When performing place matching,the sparse semantic feature maps are used to eliminate semantically movable objects,which improves the robustness of the algorithm.Also,objects with similar shapes have similar features,which makes it quick and accurate to complete the matching.Experiments on KITTI,Oxford and Mul Ran datasets are conducted,results show that the algorithm proposed in this paper has higher precision and recall rate than the existing mainstream algorithms.For the problem of outdoor mobile robot localization and mapping,a novel and universal Li DAR SLAM framework is proposed based on the research content of the previous chapters,which improves the robustness with the help of the SSFM.Based on such framework,multimetric maps based Li DAR SLAM(M3-SLAM)is constructed in this paper.It can construct both SSFM and high-precision point cloud map at the same time.Semantic information is introduced to eliminate the influence of movable objects in the steps of inter-frame registration and local map registration,which improves the accuracy and robustness of registration.Based on polar coordinate encoding point cloud context information,ring keys are constructed to extract loop candidates,and accurate loop closure detection is completed based on this.Backend optimization based on closed-loop constraints is performed to eliminate accumulated drift,and a consistent high-precision point cloud map is established based on the optimized poses.The global sparse semantic feature map and the high-precision point cloud map are used to realize a two-step coarse-to-fine global relocalization.Experiments on the KITTI Odometry dataset and self-collected dataset prove the effectiveness of the proposed method,a mean localization error of 0.83% is presented.On the basis of the above researches,multiple public datasets are used in this paper to verify and analyze the proposed algorithms.The Bulldog,an outdoor mobile robot equipped with GPS/IMU navigation system,Hesai 40-line Li DAR and other sensors,has been used as an experimental platform to conduct a large number of experiments in the campus environment.The results show that the methods in this paper have good accuracy,real-time performance and robustness.