Research On Key Technologies Of SLAM For Mobile Robots In Dynamic Scenes

Simultaneous Localization And Mapping(SLAM)is one of the core key technologies for mobile robots to realize environmental perception,decision-making and path planning.In dynamic scenes,affected by moving objects and scene changes,visual SLAM and laser SLAM All odometers have problems such as low point cloud matching accuracy,large deviation of laser SLAM pose estimation,easy to cause positioning failure and low accuracy of reconstructed maps.Aiming at the above problems,this paper discusses in depth the key technologies of visual odometry and laser odometry,the improvement method of laser SLAM pose estimation,and the effective fusion method of visual SLAM and laser SLAM.The main research contents of the thesis are as follows:(1)Research on the extraction and matching method of visual odometry feature points.In the Qtree＿ORB feature extraction part of the ORB SLAM2 visual odometry,the contrast of the image gray value is used as the o FAST adaptive threshold,and a motion estimation grid region feature matching(Motion Estimation Grid Region Feature Matching(MGR)feature matching method forms the optimization scheme of visual odometry,and rigorous comparative experiments prove that this method is helpful to improve the accuracy of visual odometry.(2)Laser odometry weighted NDT-SAICP point cloud registration method.In the dynamic scene,the high-precision point cloud registration method in the laser odometer is especially important.First,data preprocessing including ground point clustering and bilateral point cloud filtering processing is performed to obtain high-quality laser data,and a weighted NDT is proposed.-SAICP point cloud registration method,for the classic NDT algorithm,the distance weight is introduced for rough matching,and the output transformation matrix is used as the initial value of the fine matching part,combined with the SAC-IA algorithm to improve the ICP registration accuracy to complete the fine matching,comparative experiment It shows that the point cloud registration accuracy has been significantly improved.(3)Research on laser SLAM dynamic feature point removal method combined with semantic information.This paper uses a semantic segmentation network that can reach the pixel level to obtain environmental semantic information.On the basis of Deeplabv3+,an adaptive channel attention module is introduced into the encoder,and a multi-feature fusion branch structure is proposed in the decoder.Adapt to the channel attention multi-branch Deeplabv3+ semantic segmentation network model,obtain rich environmental semantic information and at the same time assign semantic masks to dynamic targets and perform dynamic feature point detection and elimination on them,and use these high-level semantic features in the environment to improve the performance of laser SLAM Positioning accuracy improves the pose estimation accuracy of laser SLAM in dynamic scenes.(4)Research on dynamic scene combined with Scan context vision and laser fusion SLAM method.Considering that there will be large inconsistencies in adjacent frames in dynamic scenes,which will reduce the accuracy of pose estimation and lead to low positioning and trajectory accuracy of SLAM,pure laser SLAM is not good at positioning in dynamic environments,and visual SLAM is good for lighting,etc.The environmental requirements are high.In order to make up for the lack of a single sensor,a vision and laser fusion SLAM method for stably extracting image features is proposed on the basis of the previous chapters,and Scan context loop detection and global pose graph optimization are added to improve the inter-frame The accumulated errors in poses form a fusion algorithm with high adaptability in dynamic scenes,and comparative experiments on public datasets show the effectiveness and reliability of this method.