Cognitive Map-Oriented Visaul Localization System

With the advancement of science and technology,autonomous driving technology has also been greatly developed,positioning is the core technology in autonomous driving.Traditional autonomous driving positioning is divided into signal-based positioning,such as GPS,base station positioning,etc.;map-based positioning,such as relying on highprecision 3D point cloud maps,uses laser radar point cloud registration.Among them,the accuracy of ordinary GPS equipment is low,the accuracy of GPS equipment using RTK technology is high,but the hardware cost is high and the positioning accuracy depends on the signal,that is,it is not applicable when there are obstacles such as tall buildings and tunnels.The cost of lidar is too high.In the prototype of autonomous driving,the cost of lidar alone is more than half of the cost of the entire vehicle.In terms of maps,highprecision 3D point cloud maps have high costs,large amounts of data,high computing resource requirements,and low update frequency.Cognitive map is a kind of self-driving map designed based on human cognitive principles.It has the advantages of small data volume,low computing resource requirements,high update frequency and relatively high accuracy.Because cognitive map design relies only on visual sensors,low cost.Cognitive map is composed of road layer,lane layer,semantic layer and dynamic information layer.Existing three-way sign-based positioning methods have strict requirements on the road environment,and it is difficult to observe three-way signs at the same time in one frame;only when the road signs are observed,positioning cannot be obtained,and vehicle position information cannot be obtained throughout;Road signs refer to the semantic information of cognitive maps.In view of the above problems,this paper proposes a visual fusion-oriented positioning method for cognitive maps,which combines the positioning results of lane lines,road signs,and visual odometers,and finally obtains the global dense autonomous driving positioning results.It has the following advantages: the cognitive map information is fully utilized,and road signs and lane lines can be used as positioning reference road signs;road conditions are low,and a single road sign can be located with a lane line.The main contributions of this article are as follows:A purely visual monocular lane line positioning model for automatic driving is proposed.Assuming that the road is a plane,the road plane parameters are pre-calibrated,and combined with the lane line perception algorithm,the position relationship of the vehicle camera coordinate system relative to the lane line can be given in the monocular situation.Get the vehicle's position relative to the current lane.A global vehicle positioning method with lane lines and single landmarks is proposed.The lane line positioning results determine the lateral position of the vehicle relative to the lane lines,and a single landmark is used to obtain the longitudinal position of the vehicle to obtain a complete and global vehicle posture.The visual odometer can infer the relative position between different frames,which is a supplement to positioning when there is no semantic landmark.By querying the cognitive map,considering the constraints of the lane line position,the global positioning constraints of road signs,and the relative position constraints of the visual odometer,the fusion optimization can obtain the global dense positioning results of the vehicle.This paper designs a set of autonomous driving positioning algorithms for cognitive maps.It is different from traditional autonomous driving positioning algorithms.It uses pure visual sensors and makes full use of environmental semantic information to obtain global position results with scale.It is one of the autonomous driving positioning algorithms.This kind of innovative exploration provides another direction for the evolution of autonomous driving technology.