LIDAR Odometry And High Precision Navigation Of Unmanned Platforms For Multi-application Scenarios

With the development of unmanned autonomous system,real-time localization and navigation researches for outdoor unmanned platforms are receiving more and more attentions.The common solution for navigation tasks of unmanned platforms at present is heavily relied on high-precision maps and GPS/INS information,but the construction of the high-precision maps needs high-precision pose information,and GPS are sometimes not available in some special environments,and the precision of GPS signal is not high enough to meet the requirements of high-precision mapping and navigation either.Therefore,the SLAM method is needed to solve the above problems,a feasible solution is to use the LIDAR to complete the localization of unmanned platforms and the construction of high-precision maps.At present,LIDAR SLAM approaches for ground applications are relatively mature,but there is no satisfactory solution for complicated,unconstructed ground/watersurface composite environments(harbors,dams,bridges,etc.).In this paper,the method of LIDAR odometry and navigation of unmanned platforms for multi-application scenarios based on LIDAR is studied in depth,and the specific contents are as follows:To solve the problem of long-distance accurate localization of unmanned platforms for multi-application scenarios,a real-time 3D LIDAR odometry(S4OM)is proposed in this paper.The odometry consists of two nodes(localization node and correction node),the localization node combines the improved Super4 PCS with the standard ICP to realize a coarse-to-fine scan matching and outputs the pose information at a high frequency(5Hz);the correction node constructs a local map with dynamic voxel grid storage structure,which can accelerate the NDT matching process between key-frames and the local map,and then corrects the localization node at a low frequency(1Hz)to obtain more accurate pose information.The S4 OM odometry system proposed in this paper can be independent of the assistance of GPS,INS and other external equipments,it can achieve good results(about 1% drift)in ground/watersurface multi-application scenarios,and has good robustness and stability.Different from 2D visual images,3D laser point clouds have the characteristics of disordered point arrangement,uneven distribution,and high noise.Therefore,there are relatively few researches on place recognition and loop detection for LIDAR at present.In this paper,a place recognition algorithm and a loop detection algorithm based on random forest model are proposed,for the problem of place recognition based on 3D point clouds,the multi-modal feature vectors composed of geometric features and histogram features are constructed to train and construct a random forest classifier,then the environmental point cloud and the map nodes are input into the classifier to complete place recognition.After that,the place recognition method is applied to the loop detection of unmanned platforms.Firstly,the loop discrimination method based on location relationship and random forest model are combined to discriminate the loops.Then geometric verification is added after the loops discrimination to obtain the real loops by point cloud overlap rate.Finally,the experiments are carried out to prove that the proposed method has good accuracy and robustness.On the basis of the above researches,a 3D LIDAR SLAM algorithm(S4-SLAM)based on graph optimization framework is constructed by combining the S4 OM odometry with the loop detection method based on random forest model.The algorithm consists of front-end odometry(S4OM),loop detection and back-end optimization.The odometry can output the pose information of the unmanned platform at a high frequency,and the key-frames of pose graphs are acquired at the frequency of 1Hz,when the loop detection algorithm detects that the trajectory of the unmanned platform constitutes a loop,the global graph optimization is performed,which makes the trajectory of the unmanned platform closer to the real trajectory and improves the mapping precision.Finally,the place recognition method and NDT matching algorithm are combined to achieve the global localization of unmanned platforms in the high-precision map.The experimental results show that the high-precision map constructed in this paper is feasible in the application of unmanned platform navigation.On the basis of the above studies,the KITTI dataset is used for experimental analysis and verification,and the evaluation is carried out on KITTI Odometry website.Besides,the wheeled unmanned vehicle and “Jinghai no.3” unmanned ship with multiple sensors like 3D LIDAR,IMU and GPS are used as the experimental platforms to do a wide range of localization and navigation experiments in outdoor multi-scene environment(harbour,campus),to prove the effectiveness and correctness of the proposed methods in this thesis.