Research On 3D Mapping Of Smart Car Based On Machine Vision

Entering the 21st century,with the arrival of the fourth industrial revolution,smart vehicles and driverless technologies have attracted widespread attention from academia and industry as a new generation of travel methods.For smart vehicles to realize unmanned driving,first of all,it should be able to get a map of its environment and its position on the map,SLAM（Simultaneous Location and Mapping）technology is proposed to solve this problem.Among them,the threedimensional mapping of visual SLAM technology based on machine vision,which has the characteristics of rich environmental information,relatively low cost,and convenient installation,has gradually become the mainstream development direction of SLAM technology.However,the traditional VSLAM algorithm cannot meet the needs of autonomous navigation and obstacle avoidance of smart cars,the sparse landmark map built by it has defects such as uneven feature extraction,high false matching rate,and weak robustness.Therefore,this paper designs a 3D map construction system based on the traditional ORB-SLAM2 algorithm.The main work content is as follows:（1）This paper first introduces the basic theoretical knowledge of the visual SLAM system and the principle of depth camera imaging,and calibrates the Kinect V2 camera used in this paper;（2）Aiming at the overlap problem of the original ORB-SLAM2 algorithm visual odometer feature extraction,an improved ORB-SLAM2 visual odometer is proposed.By meshing the image pyramid,FAST corner point adaptive threshold extraction is completed to eliminate the overlapping phenomenon of feature extraction and make the distribution more uniform;for feature mismatch phenomenon,use RANSAC（random sampling agreement）algorithm to filter out,and finally use the ICP（Iterative Closest Points）method to complete the pose estimation.（3）In view of the lack of spatial structure information of the sparse landmark map and the inability to realize the positioning function,a 3D point cloud map construction method based on RGB-D camera is proposed,which can realize dense map construction;Subsequently,in order to solve the defect of the point cloud map running with a large load,on the basis of it,a mapping method based on Octree was designed.Through comparative experiments,the storage space occupied by the octree map at different resolutions has decreased by 85.4% on average;in addition,when the algorithm is running,CPU load rate dropped by 2.6% on average,and memory usage dropped by 64.1% on average,proving that it has better compression performance and real-time performance.（4）Compare the positioning accuracy of the smart car 3D mapping algorithm designed in this paper with the original ORB-SLAM2 algorithm;first,use the four sequences in the public data set TUM for testing,compared with the original ORB-SLAM2 algorithm,the ATE（absolute pose error）of this algorithm has dropped by 36.95% on average,and the RPE（relative pose error）has dropped by 23.5% on average;then the comparison experiment with the smart car in the real indoor environment is implementing,the trajectory of the algorithm in this paper is smoother,and there is no obvious positioning drift and frame loss,the positioning accuracy is significantly improved;the experimental results show that the smart car 3D mapping algorithm proposed in this paper has a higher feasibility and robustness.