Multi-sensor Fusion Based High Precision 3D SLAM Method For Large-scale Dynamic Scenes

Li DAR and Inertial Measurement Unit(IMU)fusion based simultaneous localization and mapping(SLAM)method becomes an active research topic in the field of self-driving,since it is robust against illumination variation,and it achieves high accuracy of pose estimation and consistency of mapping in various environments.At present,most fusion based SLAM methods are implemented under the assumption that the scene is static and the sensor is moving at low speed.However,there are moving objects in the real scene,and the sensor moves at high speed.In complex and changeable dynamic scenes,eliminating the distortion caused by the sensor's self motion and achieve accurate localization and robust mapping is a problem to be solved by SLAM for large-scale dynamic scenes.To solve this problem,based on the existing fusion framework and the characteristics of large-scale dynamic scenes,this paper proposes a multi-factor graph optimization based SLAM method.Based on scene segmentation,we achieve reliable multi-object detection and tracking,and propose improved mtehods on back-end optimization,loop closure detection and other key modules,which improves the accurate of localization and consistency of mapping for large-scale dynamic scenes.The work done in this paper mainly includes the following aspects:(1)A multi-object tracking method based on point cloud segmentation is proposed.A Fully Convolutional Nerual Network(FCNN)based point cloud segmentation method is used to detect potential moving objects in the point cloud.Then,the Unscented Kalman Filter(UKF)is used to optimize the parameters of the estimated moving objects,and the Probabilistic Data Association(PDA)is further used to constrain the update of the prediction equation.Based on this,Interactive Multi-model(IMM)method is used to combine the state estimation results under different models.Finally,the method achieves reliable multi-object tracking and outputs speed,category,and status flags of the moving objects.(2)A factor graph optimization based simultaneous object tracking and SLAM method is proposed.Based on the results of multi-object tracking,the scene is divided into a static part and a dynamic part.For the static part of the point cloud,factor graph optimization based SLAM method is used to achieve pose estimation and mapping.The front-end manages key frames and achieves the data association of key frames with sliding windows and the back-end optimizes different types of information based on the graph optimization model,which achieves accurate pose estimation and mapping for dynamic scenes.(3)An improved multi-factor graph optimization based high-precision SLAM method is proposed.We combine IMU pre-integration and factor graph optimization to generate IMU factor.A point cloud descriptors based loop closure detection method is used to achieve accurate and efficient loop factor generation.The proposed method improves the existing multi-factor alignment method and achieves accurate alignment of several pairs of factors through interpolation,andthe degradation process when the sensor fails is discussed,which improves the existing multi-factor graph optimization based SLAM method.(4)In the Linux environment,four methods are experimentally verified based on the Robot Operation System(ROS).Experiments on multi-object detection and tracking method are carried out on the object subset of the KITTI dataset,which prove that the proposed method can achieve reliable multi-object detection and tracking in dynamic scenes.Quantitative and qualitative experiments are carried out on three datasets,which prove that the simultaneous objects tracking and SLAM method can achieve accurate pose estimation and build consistent static maps in dynamic scenes.The loop closure detection method is tested on the complex sequence of the KITTI data set,which proved that the method can achieve stable loop closure detection in reverse access scenes.The improved multi-factor graph optimization based SLAM method and proposed loop closure detection method conduct multi sets of comparative experiments on two datasets,which prove that the proposed method can achieve accurate pose estimation and consistent mapping for large-scale dynamic scenes.48 Figures,11 Tables,and 66 References.