| Recently,with the gradual development of AI(Aritificial Intelligence)technology,the entire automotive industry has paid more attention on intelligent application and development,especially the driverless field.Many Internet giants and innovative companies have invested in this field.They have invested lots in this field and tried to carry out the intelligent and large-scale trial of driverless in various svenarios.Generally,navigation system plays an important role in this field.In addition to providing accurate ego-car observations,it also needs to be combined with other systems to construct the HD map,and provide basic input for real-time planning and decision.In this system,two more important types of problems are Ego-Car state estimation and Simultaneous Localization and Mapping(SLAM).With the deepening of theoretical research and practice,the academic circles generally believe that the multi-sensor fusion can better improve the accuracy and robustness of the system,which is an important development direction in recent years.This paper also studies the schemes and methods of ego-car state estimation and mapping based on multi-sensor fusion.The main contents are as follows:Firstly,combined with the camera and Lidar data,the Seg-Fusion model is proposed based on Yolo V3 object detector and ground removal algorithm.It can be used to segment the obstacle point clusters directly from 3D point cloud,to achieve the perception part.Secondly,We used the IMU pre-integration method,to help design our Camera-Lidar-IMU loosely coupling system;The ground points are used to process Ground Normal Vector Registration,and the obstacle points are used for Iterative Closest Point Registration(ICP);Then their results are fused to achieve odometry in out-door scenes with sparse features.Thirdly,based on the Iterative Extended Kalman Filter,the Multi-Sensor fusion system is proposed for ego-car state estimation.The system uses a parallel pipeline of different sensors,to use data to achieve different pre-process algorithm;this scheme improve robustness of the system.In the experiment,the failure simulation of the system was carried out,and the output of the system remained stable when different sensors were failed.Finally,the Iterative Error-State EKF(i ESKF)is proposed,and the Multi-Sensor Fusion Lidar Odometry and Mapping(MSF-LOAM)system is designed based on the i ESKF.The i ESKF combines the advantages of both ESKF and i EKF,can approximate the optimal solution through iteration;and it is calculated near the state space,which can avoid system issues caused by measurements.By fusing the GPS,our system can also apply for out-door scenes with sparse feature.Compared with main-stream LOAM systems,the MSF-LOAM has a smaller cumulative error and higher accuracy. |
