Research On Multi-level Cognitive Map Expression And Generation Method For Intelligent Vehicles

With the development of technology,smart cars are gradually emerging,and maps play an increasingly important role as an important part of ensuring the safe,stable and efficient operation of smart cars.Since the vehicle does not have the ability of humans to highly analyze environmental information,traditional maps are difficult to meet the needs of smart car operation,and high-precision maps for smart cars have emerged.Highprecision maps have high precision in modeling the environment and provide richer and more accurate environmental information than traditional maps.The current popular smart car map scheme is built by means of lidar-based multi-sensor fusion,and its map construction cost is high,the update speed is slow,and the use cost is high.In view of the above problems,this paper proposes a multi-level cognitive map expression and generation method for smart cars.This method is mainly based on visual information,perceives semantic information in the road environment,and constructs a cognitive map of smart cars.It has the characteristics of low hardware cost,small amount of calculation,and fast map update.The main contributions of this paper are as follows:A visual positioning and mapping method for joint point line features is proposed.In the outdoor scene,due to the characteristics of illumination and seasonal changes,the traditional feature point construction map method is not robust enough.Adding environmental line segment feature constraints improves the accuracy of visual map construction.The method extracts image point and line segment information,and combines the point line feature to perform image tracking.After generating the key frame,the local map is constructed,the point cloud is generated,and the map is optimized by the principle of map optimization to obtain a visual point cloud map.A cognitive map generation method for smart cars is proposed.When the smart car is running,the visual point cloud map is constructed.Because there are other moving vehicles and pedestrians on the road,the visual tracking will reduce the accuracy of the visual point cloud map.Therefore,using deep learning semantic perception,detecting candidate dynamic targets and judging whether they are moving targets through moving consistency,filtering out dynamic feature points and performing visual tracking to construct a visual point cloud map.The visual map based on image tracking will produce image accumulation error,and the GNSS high-precision positioning will also deteriorate the positioning accuracy in some areas(occlusion,etc.).Therefore,an image pose estimation method based on fusion visual positioning and GNSS high-precision positioning information is proposed to construct high.Precision visual map.Highprecision visual map combined with semantic perception,extracts road semantic features,identifies and locates them,and finally generates cognitive maps.This paper designs a vision-based smart car cognitive map system to meet the needs of smart car operation.The road environment is modeled by hierarchical map expression,including road layer,lane layer,semantic layer and dynamic information layer,which can express the intelligent driving scene well.This method can not only meet the characteristics of efficient storage and efficient use of map data,but also facilitate the update and expansion of later maps.The method is aimed at the smart car map data volume,high application cost and slow update,and realizes the way of constructing map based on visual information,which provides support for safe,stable and efficient operation of smart cars,and has made a problem for smart car map construction.Exploratory research.