Study On The Key SLAM Technologies Of Mobiel Robot Based On Lidar And Vision

With the development of CPU and price fall of new sensors in recent years,the artificial intelligence industry has developed rapidly.Mobile robot is the product of many interdisciplinary studies,which include localization,environment detection,human-computer interaction,motion control,path planning and so on.Accurate localization and environment detection are the most important and fundamental technologies.But mobile robot is being restricted by fixed application scenario,unreliable localization result and weak environment interaction ability,so the mobile robot system with high intelligence and high reliability cannot be realized.Based on the above requirements,this thesis is intended to study the key technologies in multi-sensor calibration,3D-2D/3D-3D pose calculation algorithm,pose graph optimization,tight combination of lidar and camera.Based on the innovation or improvement of some algorithms and models,the accuracy of localization and mapping has been improved.Details of research study are as follows:1.Camera intrinsic and camera-lidar extrinsic parameters calibration algorithm based on trihedral are proposed.By analyzing the disadvantages of state-of-art camera and lidar calibration algorithms,a new camera intrinsic parameter and camera-lidar extrinsic parameter calibration algorithm,which is based on trihedral calibration plate,are put forward.When planar calibration plate is used,this results in inaccurate camera pose in calibration process.Compared with planar plate,trihedral calibration plate guarantees that all of 3D points are not on the same plane,so it can improve the accuracy of camera poses and intrinsic parameter.Camera-lidar calibration also can use this trihedral calibration plate to obtain extrinsic parameters.This calibration algorithm does not need complicated stage and can obtain camera intrinsic and lidar-camera extrinsic parameter together.2.Improved PnP algorithm based on PCA framework is proposed.State-of-art PnP algorithms can not obtain accurate and robust results when variance of input 3D points is very large.A robust PnP algorithm framework,that is based on PCA method,is developed in this thesis.This robust framework uses PCA method to establish a new coordinate and conduct PnP calculation in the new coordinate.Then a new pose transformation model is applied to this framework to improve the accuracy of translation result.Input 3D points of traditional PnP experiment using simulated data are in narrow range and cannot reflect the real situation.In this thesis,the simulated experiment uses actual data to set simulated data to make up traditional experiments.At the same time,the actual data is also used to test the robustness of PnP algorithm.According to experiments,our method is more accurate and robust than state-of-art algorithms.3.Total least squares of ICP algorithm based on Lie Group is proposed.Common-used 3D-3D(ICP)algorithm is the approximation of the input data error model and cannot describe model accurately.Total least squares ICP algorithm based on Lie algebra is put forward to solve this problem.In this model,the errors of input source and target data are taken into consideration.Compared with traditional TLS ICP algorithm,Lie algebra is used to replace with Euler angle to solve gimbal lock problem.Two different parameterization models based on Lie algebra are put forward and give the analysis of these two models.In the experiments,TLS-ICP based on Lie algebra algorithm is compared with LS-ICP algorithm using actual and simulated data.4.Incremental pose graph optimization algorithm which does not rely on FEJ is proposed.By analyzing the drawback of incremental PGO algorithm,a new incremental PGO algorithm(G-PGO)which is not dependent on FEJ is invented.Compared with existing incremental PGO algorithm,the new algorithm guarantees that the converge direction is towards global optimal solution.Current PGO initial technique and cost function is studied in this thesis and choose the best performance one to be applied in G-PGO algorithm.A loop closure outlier detection error metric is used in G-PGO algorithm to exclude outliers in optimization process.Experiments are given to prove that G-PGO algorithm is more accurate than current incremental PGO algorithm.5.Tight combination of lidar and camera based on plane is proposed.The method of tight combination of lidar and camera is described.Plane information is used to realize localization and mapping.Firstly,a geometric meaningful curvature is invented by taking sparse 3D lidar measurements into consideration.Secondly,plane detection method which is based on three-times region growing is described in details.In this method,we convert XYZ-space to voxel space to conduct BFS search.Thirdly,a polygon edge point detection method is invented based on dilation and erosion,which is used to conduct plane match.Lastly,a pose optimization method which does not use map point information is designed to recover plane dense cloud in space.