| With the continuous development of science and people’s aspiration for a better and convenient life,autonomous driving technology has gained widespread attention.The accuracy of positioning results will directly affect the results of path planning and navigation in autonomous driving technology.In the process of localization and map building,cameras can obtain good environmental image information,but are not suitable for environments with low texture and light variations.Lidar,on the other hand,is rich in geometric structure information,so fusion based on camera and multiline lidar has good complementarity.In this paper,we propose a fusion localization and map building algorithm based on camera and multi-line lidar to enhance the robustness of the whole system and the accuracy of results,the main research of this paper is as follows:First,preprocessing of visual image data.A method for evaluating the effect of feature point extraction and tracking algorithms is proposed,which is to compare the results of KLT tracking with four mainstream feature point extraction algorithms: ShiTomasi,ORB,SIFT,and SURF,and selecting the best set of methods according to the experimental results,and homogenizing the feature points to avoid too concentrated feature points affecting the tracking and optimization.The LSD algorithm is used to extract the line features from the image,and an improved algorithm for line feature screening is proposed,i.e.,line features of similar descriptions are removed from the image to reduce line feature mismatch and improve the computational efficiency.The use of lidar-assisted visual initialization,combined with point cloud information,can effectively avoid the defects of scale loss in the initialization process of vision.Then,preprocess the lidar point cloud data.Completed point cloud downsampling,ground point cloud segmentation,and object segmentation processing,which can reduce the amount of post-processing data while retaining the useful features of the point cloud.Then,on this basis,the edge point features and plane point features in the point cloud are extracted,which are used for inter-frame matching of the point cloud and optimized poses as constraints.Finally,in order to obtain better localization and mapping results,a visual lidar fusion algorithm based on sliding window visual point-line features and point cloud feature constraints is proposed,i.e.,a cost function is established based on the relationship between the visual key frames of point and line features and the relationship between the corner and plane features of point cloud frames.Experiments were conducted using the KITTI dataset to compare the localization and mapping results of Le GO＿LOAM,pure Li DAR,and visual-Li DAR fusion algorithms.The results show that,in terms of root mean square error,the visual-Li DAR fusion algorithm reduces errors by 61.9% and 97.8% compared to Le GO＿LOAM and pure Li DAR algorithms,respectively,on sequence 00;and reduces errors by 44.6% and91.7% on sequence 05.This indicates that the visual-Li DAR fusion algorithm performs much better than the pure Li DAR algorithm in terms of localization and mapping,effectively compensating for the deficiencies of a single sensor,significantly improving the accuracy and robustness of the algorithm,and outperforming the Le GO＿LOAM fusion algorithm. |
