Lane Marking Recognition Algorithm With Autonomous Evolution Ability Based On UAV

High definition map is an important part for building smart cities,which can effectively help driverless,map navigation,road condition analysis and many other practical application scenarios.At present,the manual construction of high definition map,which is widely used in the industry,reveals the problems of huge map fabricating cost and low timeliness efficiency in the face of problems and requirements such as massive highway scenes and high-frequency map updating.Therefore,a method that can automatically build high definition map quickly and accurately highlights its importance.The rise of artificial intelligence algorithm in the field of image processing makes it possible to use intelligent means to quickly build high definition map.At the same time,the development of UAV hardware equipment provide a new alternative way for large-scale and real-time data acquisition in smart cities.With the help of UAV hardware platform,using artificial intelligence algorithm and traditional image graphics algorithm,this paper realizes the complete process from lane marking image acquisition to automatic construction of high definition map,speeds up the data flow link from image data to high definition map data,and foster the applications of smart cities.The lane markings in the UAV perspective image show diversified shape characteristics,such as straight line,bifurcation line,intersection line,helix and so on,which poses a certain challenge for accurately describing the shape of lane markings in artificial intelligence algorithm.In this paper,Bi Se Net V2 real-time semantic segmentation model is used to complete lane marking recognition,and pixel points are used to describe the shape of lane marking,which can represent the lane marking of various line shapes with pixel accuracy,and meet the requirements of UAV onboard computer environment in terms of model processing timeliness.In addition to the complex shape of lane markings,the pixel proportion of lane markings in UAV perspective image is much lower than that of background pixels,which belongs to the problem of imbalance data.Therefore,this paper proposes an unbalanced joint loss function of samples based on Top-k,so that the model can learn the information of each sample category evenly in the training process.In addition,according to the characteristics of the scene in which the UAV is flying,this paper designs image data enhancement algorithms such as random shape generation and random image defogging,so as to increase the data feature richness of the image data set.Considering the data island problem of model training,this paper proposes an off-line semi autonomous update mechanism of data set,which combines the algorithmic filtering and manual filtering of labeled data to realize fast image labeling and build a closed loop between data set and model.At the same time,considering the different emphasis of difficult and easy samples in long-time model training,this paper proposes a single task sample data adaptive resampling mechanism to achieve the balance between difficult and easy samples in the training process.Under the hardware environment of NVIDIA Jetson NX,the lane marking recognition algorithm proposed in this paper can achieve the 0.9920 pixels accuracy on 640 ?640resolution image and the image processing speed is 5～10 frames per second.After lane marking recognition,this paper uses Zhang Suen image thinning improved by edge filling,lane marking instance identification algorithm,lane marking vector extraction algorithm based on connected components analysis,linear vector aggregation algorithm based on gradient and position,etc.to realize the conversion from lane marking recognition results to Lane Let2 static high definition map format.After the field test of expressway,The map accuracy of 1.8～3.5 meters can be achieved on a single control point of lane marking.