Construction Of Semantic Mapping Based On Laser SLAM And Binocular Camera

Conventional laser SLAM technology has the problem of cumulative localization error due to its single sensor,and the constructed environment map contains only geometric information,which is not sufficient to recognize the high-level semantic information of the surrounding environment.In order to better apply to practical scenarios,this paper fuses binocular and laser information,and designs a semantic map construction method based on image semantic segmentation and SLAM technology.This paper improves the localization accuracy by fusing data and constructs a point cloud map with semantic information to achieve the understanding of the surrounding environment and objects.The main work of the project contains the following aspects.Firstly,two kinds of odometer nodes in laser SLAM and visual SLAM are studied.The working process of the algorithm is analyzed to pave the way for the subsequent data fusion to improve the positioning accuracy.Secondly,for the problems of incomplete segmentation and low accuracy of Deep Lab V3+semantic segmentation algorithm for small targets,an improved Deep Lab V3+ algorithm is proposed in this paper.Combined with the idea of structural reparameterization,part of the structure of the network is reconstructed using the DBB module,and the network is trained using a hybrid loss function consisting of cross-entropy loss function and Dice loss function to further improve the accuracy of segmentation and refine the problem of poor segmentation of small targets.The proposed algorithm is validated on the Cityscapes datasets,obtained MIo U values is 76.79% and MPA is 83.95%,which is 4.63% and 3.53% higher than the original algorithm,respectively,confirming the effectiveness of the improved algorithm.Then,this paper designs a semantic map construction method framework based on LOAM algorithm and Deep Lab V3+ semantic segmentation algorithm.By determining the spatial position conversion relationship between the laser radar and the camera,the three-dimensional point cloud is projected to the pixel coordinate system to achieve the purpose of mutual mapping between the laser point cloud and camera image data.The improved Deep Lab V3+algorithm is used as the semantic information extraction module to process the image input by camera and output the semantic segmentation map,and the corresponding semantic attributes are obtained from the 3D point cloud through projection,and then the semantic point cloud is published to the map building node in the LOAM algorithm to complete the construction of semantic map.In addition,a combined positioning method is designed based on the extended Kalman filter algorithm.By fusing the data of lidar and binocular visual odometry,the optimal estimation of pose is caried out to improve the cumulative error of LOAM algorithm due to its single laser sensor and increase the positioning accuracy.Finally,the proposed algorithm is programmed and implemented in this paper to complete the whole algorithm process.The proposed algorithm is experimented and analyzed with the standard dataset KITTI,and the experimental results show that the semantic map construction method proposed in this paper can improve the localization accuracy and effectively construct point cloud map with semantic information.