Study On Simultaneous Location And Mapping Of Intelligent Vehicles Based On Lidar Point Clouds Data

The environmental perception of intelligent vehicles aims to detect and recognize obstacles in the surrounding environment of the vehicle,track dynamic objects,and map static objects.It is the basis for intelligent vehicles’ decision-making,planning and motion control.Simultaneous Localization and Mapping(SLAM)is one of the core technologies of environmental perception.It means that in the absence of any prior information,the vehicle only relies on its own sensors to perceive and map the surrounding environment,and to locate its own position in the built map.There are many forms of map established by the intelligent vehicles through sensors,such as occupancy grid map,feature-based map,and topological map.Among them,the feature-based map has received extensive attention and successful application due to its advantages of less memory occupation and intuitive presentation environment.At present,the SLAM method based on point features must adopt a data association algorithm to establish a corresponding relationship between the target state and the measurement information.However,the SLAM algorithm based on data association is less robust in high-clutter scenes,which easily leads to incorrect estimation of features.In view of this,this paper adopts the SLAM algorithm based on Random Finite Sets(RFS),which avoids the complicated data association process,and has high robustness even in high clutter scenes.This paper carries out in-depth research from the following aspects:This paper first introduces the basic knowledge of RFS and the basic principles of RFS-based filters.On the basis of using RFS to model the SLAM problem,the recursive formula of RFS-SLAM under the Bayesian framework is given,which provides a theoretical basis for the specific application in the following.Aiming at the problem of the large amount of data in Lidar sensors and the long time-consuming direct feature extraction,this paper proposes a set of point clouds data processing procedures.In the stage of preprocessing,point clouds down-sampling algorithm,pass through filter,radius filter and ground point clouds removal algorithm are used to effectively reduce the number of points.In the stage of point clouds segmentation,a density-based clustering algorithm is used to segment the preprocessed point clouds,and points with similar attributes are clustered into clusters.In the step of feature extraction,the model fitting methods MSAC(M-estimator Sample Consensus)and MSSE(Modified Selective Statistical Estimator)are used to extract columnar features and planar features in the environment respectively,and obtain perfect extraction performance.Then,in view of the problem that the data association algorithm is prone to errors in the high-clutter environment,which affects the accuracy of environmental feature estimation,this paper uses RFS to model the map features and sensor measurements,avoiding the complicated data association process.The proposed RB-PHD-SLAM algorithm uses a particle filter to transfer the probability density of the vehicle trajectory on the one hand,and on the other hand uses a PHD filter based on sequential Monte Carlo to transfer the posterior intensity of the map.Aiming at the problem that it is difficult to predict the location of the new-born target in the autonomous driving scene,this paper uses the measurement information at the previous moment to predict the location of the new-born feature.The effectiveness of the proposed RB-PHD-SLAM algorithm is verified by simulation data and test datasets.The verification results show that the RBPHD-SLAM algorithm is superior to the traditional random vector-based Fast SALM algorithm in terms of vehicle localization accuracy and environmental feature estimation accuracy.Finally,in order to provide more intuitive environmental information,this paper additionally uses planar features to map the environment.Aiming at the problem that the planar features extracted by the MSSE algorithm have no boundary information,the plane segmentation is realized by the method of establishing a search box,and the boundary information of the planar features is obtained.In addition,this paper uses three threshold conditions to determine whether two features are derived from the same structure,and merges the features that meet the conditions to remove a large number of redundant features in the global map.The KITTI dataset is used to verify the mapping method,and the results show that the method can realize the map construction based on planar features in a structured environment.