Research On The Optimization Method Of Visual Odometry Based On Remote Sensing Images

With the rapid development of artificial intelligence,driverless technology has received extensive attention and has become a research hotspot in the field of artificial intelligence and automobiles.The key technologies of unmanned driving mainly include modules such as environmental perception understanding,environmental modeling,decision-making and control.Positioning technology is an important part of environmental perception.Accurate positioning is the basis for autonomous navigation of unmanned systems,and is the prerequisite for subsequent environmental understanding and modeling,and path planning and navigation.Traditional positioning methods mainly use GPS,IMU and other devices.Currently,there are three main types of visual positioning: one is the simultaneous positioning and mapping by SLAM technology,the other is the global positioning based on existing maps,and one is the visual odometry based on local pose estimation.This paper mainly studies and analyzes the unavoidable cumulative error in the visual odometry,and proposes a scheme of using remote sensing images as prior information to assist visual positioning.The method proposed in this paper is mainly aimed at the unknown scenes that the unmanned system cannot enter in advance.We use the current remote sensing or drone technology to shoot the target area and extract the road to assist the visual odometry positioning.Compared with the traditional GPS,the IMU positioning method has the advantages of being less constrained by the scene,more flexible,lower cost,and capable of reducing accumulated errors.Compared with the current visual positioning method,it has the advantages of being independent of the scene,eliminating the need to enter the positioning scene for mapping,reducing accumulated errors,and surveying unknown scenes.The vision positioning method of unmanned system based on remote sensing images proposed in this paper is to extract roads from remote sensing images and construct a road map.The road map is used to correct the cumulative error of the visual odometry and correct the trajectory.The main research contents of this article include:1)An end-to-end road extraction method based on neural networks is proposed and implemented,and it is proved by experiments that it can accurately segment roads from remote sensing images.2)Based on roads extracted from remote sensing images,this paper uses skeleton extraction,intersection extraction,and coordinate transformation to build a road map.The road map consists of a road network map and an intersection topology.3)This paper proposes a way to describe road features,that is,the change in road direction.At the same time,a method for extracting candidate points in road maps is proposed,which aims to describe road features through candidate points.At the same time,the trajectory representation method and similarity measurement method proposed in this paper can accurately perform trajectory matching.4)This article uses visual odometry to describe the unmanned system motion trajectory,based on the road map and trajectory matching algorithm to correct the cumulative error in real time,and correct the pose for the candidate points in the road map.5)This article uses the matching of the motion trajectory and the candidate points to return the current position to the road map,which can restore the exercise trajectory of the unmanned system in the road map.Finally,this paper proves the effectiveness and flexibility of our positioning scheme through a large number of experiments,and uses this scheme on the driverless platform of the project team to prove the feasibility in the actual environment.