Research On Key Technologies Of SLAM Based On LiDAR And GNSS

As one of the key technologies to realize robot autonomy and intelligence,Simultaneous Localization and Mapping(SLAM)technology has been widely used in space exploration,unmanned driving and service robots.LiDAR(Light Detection and Ranging)has become one of the main sensors used in SLAM technology due to its long range,immunity from light interference,and high measurement accuracy.However,the expanding application scope and the increasing complexity of scenes have left many shortcomings in SLAM technology in terms of localization accuracy,real-time performance,environmental robustness,and dynamic scene adaptation.At the same time,the high-precision and absolute localization information output by Global Navigation Satellite System(GNSS)can effectively provide compensation for LiDAR SLAM when the scene structure degrades.The paper aims to improve the accuracy,robustness,and real-time performance of LiDAR-based SLAM.To address this,both LiDAR-based and LiDAR/GNSS fusion-based SLAM algorithms are studied.The paper takes the Iterative Closest Points(ICP)-based LiDAR SLAM algorithm as the starting point,and improves the localization and mapping performance of traditional algorithms in scenes with sparse features by the point cloud segmentation.On this basis,a LiDAR SLAM algorithm based on higher-level semantic information is further proposed,which improves the localization and mapping performance of traditional algorithms in structured scenes with dense features.Furthermore,considering the accumulative drift of the LiDAR SLAM algorithm and the dependence of traditional fusion algorithm on the sensor model,a SLAM solution based on curve deformation for combining LiDAR and GNSS is proposed.Finally,aiming at the problem of the above fusion algorithm in the scene with unstable GNSS signals,an improved LiDAR/GNSS fusion SLAM algorithm is further proposed.The specific research includes the following four aspects:(1)Aiming at the problem that the traditional ICP-based LiDAR SLAM solution has low corresponding points search efficiency in sparsely-featured scenes and the dynamic points in the environment are likely to cause mismatching of corresponding points,a LiDAR SLAM algorithm based on point cloud segmentation is proposed.The large number of ground points in the point cloud not only reduces the efficiency of the point-to-point nearest neighbor search in the ICP algorithm,but also interferes with the corresponding point search that is only constrained by the distance between the points.To this end,a ground point segmentation and down-sampling method combining scan line and depth image is proposed to improve the efficiency of the algorithm and the accuracy of pose estimation.Furthermore,aiming the problems that the noise points and dynamic points in the environment will cause mismatching of corresponding points,a method based on point cloud segmentation for removing noise points and dynamic points is proposed.On this basis,a back-end pose graph optimization framework is introduced to further correct the accumulated drift caused by ICP.The experimental results show that the proposed algorithm outperforms the traditional ICP-based and state-of-the-art features-based LiDAR SLAM solutions in sparsely-featured scenes in terms of localization and mapping accuracy and efficiency.(2)Aiming at the problem that traditional SLAM algorithm based on geometric features namely,LiDAR Odometry and Mapping in Real-time(LOAM),suffers from inaccurate feature extraction and corresponding point matching in structured scenes with dense features,a LiDAR SLAM solution based on semantic information is proposed.Considering that dynamic objects in the environment,such as high-speed vehicles,can cause incorrect data association which can lead to estimation errors,this paper first proposes a dynamic objects filtering strategy by combining geometric and semantic information.Then,aiming at the problem that the local smoothing-based feature point extraction method cannot guarantee accurate feature extraction,a semantic labels-based feature extraction method is proposed.In addition,feature-based corresponding point matching is essentially a nearest neighbor search based on feature point constraints,which cannot adapt to complex and changing environments.Therefore,a corresponding point search method based on semantic label constraints is proposed.The experimental results show that the localization accuracy and mapping performance of the proposed algorithm in densely structured scenes are better than traditional ICP-based and feature-based LiDAR SLAM algorithms.(3)Aiming at the problem that traditional LiDAR SLAM algorithms suffer from accumulation drift and filtering-based fusion methods rely on sensor models,this paper proposes a LiDAR/GNSS fusion SLAM algorithm based on curve deformation for the scene with stable GNSS signal.Traditional LiDAR/GNSS-based fusion methods mostly use filtering frameworks such as Extended Kalman Filter(EKF)and Particle Filtering(PF).However,EKF is susceptible to linearization errors while PF suffers from particle degeneracy.Different from the filtering-based fusion methods,this paper proposes a LiDAR/GNSS fusion method based on curve deformation from the perspective of computer graphics.In addition,considering that the above fusion algorithm outputs 3-degree-of-freedom rather than 6-degree-of-freedom poses and the point cloud distortion will cause estimation errors,the fused poses and 3D LiDAR are further fed into a semi-rigid SLAM back-end optimization framework to obtain accurate localization and mapping results.The experimental results show that the proposed method is superior to the one before improvement in terms of accuracy.(4)Aiming at the problem that the existing SLAM algorithm based on LiDAR/GNSS fusion has insufficient scene adaptability,low accuracy and real-time performanc in the scene with unstable GNSS signals,an improved LiDAR/GNSS fusion SLAM algorithm is proposed.The framework consists of three parts:feature extraction,LiDAR odometry and LiDAR mapping.For feature extraction,aiming at the problem that the traditional feature extraction method based on local smoothness cannot accurate extract features,this paper proposes a vector angle-based feature point extraction method.Aiming at the accumulative drift error of the LiDAR odometry,a GNSS-assisted LiDAR odometry method is proposed.At the same time,when GNSS participates in fusion localization,real-time calculation of GNSS localization accuracy is required,so that the fusion algorithm can adjust the weight of GNSS in the fusion algorithm in real time according to the accuracy of GNSS.Existing LiDAR/GNSS fusion schemes often use fixed weighting factors,or use Position Dilution of Precision(PDOP)to qualitatively determine the output accuracy of GNSS.However,this method suffers from information delay and large errors.To this end,a GNSS precision factor calculation method based on laser point cloud is proposed.Finally,in the LiDAR mapping module,a framework for integrating GNSS localization information into the LiDAR mapping process is proposed by combining the obtained GNSS precision factor and the improved curve deformation-based fusion algorithm in(3).Experimental results show that the proposed algorithm has better position estimation accuracy,real-time performance and robustness in the scene with unstable GNSS signals than the previous algorithm and the typical fusion algorithm.Finally,the experimental data of different scenarios verify the effectiveness of the algorithms of each chapter in different application scenarios.