Research On Track Control Of Road Roller Based On Deep Learning

With the development of deep learning,the integration of deep learning and driverless technology is also deepening.The realization of unmanned operation of roller can improve operation efficiency and compaction quality.The traditional method provides a navigation benchmark for the unmanned operation of the roller through machine vision technology.However,it is difficult to extract feature points and the process is complex in pavement compaction.In order to solve the above problems and realize the unmanned operation of the roller,the deep learning technology is applied to the roller.Based on the deep learning semantic segmentation network technology,this paper extracts and processes the edge features of the road surface to be compacted,and studies the operation trajectory control of the roller.However,the frequency of use of road rollers is much lower than that of automobiles.It is difficult to collect pictures of road compaction scenes of road rollers,which leads to the lack of special data sets for road compaction scenes of road rollers at present.It is difficult to obtain data sets of the same scale as automatic driving of automobiles to ensure the training quality of the network by relying on pictures taken by road rollers during construction.However,there are many and relatively perfect data sets for unmanned driving of automobiles.Therefore,the vehicle driverless data set can be used for the training of the roller semantic segmentation network by the method of data set migration.This paper analyzes the vibration difference between the roller and the car itself and the different characteristics of the road surface materials identified,which lays the foundation for the data set migration.Firstly,the difference between the vibration of road rollers and automobiles may affect the segmentation accuracy of the road edge by semantic segmentation network,thus affecting the effect of data set migration.8000 road image data were collected for semantic segmentation experiment on Deeplabv3+ network after training and optimization.Under different vibration conditions of roller on and off steel wheel and normal car running,Io U values of pavement area segmentation were 92.039%,95.887% and 93.208%.Finally,the experimental data prove that the vibration difference has no effect on the migration of the data setSecondly,there are not only asphalt pavement but also soil pavement in the road compaction scene.It is only confirmed that vibration has no influence on data set migration,which is not enough to fully explain that the automobile data set can be used for road rollers.In the compaction scene,it is more difficult to acquire the image of soil pavement and establish the data set,and the automobile data set is basically asphalt pavement image,so the method of feature transfer can be used to generate soil pavement data based on the automobile asphalt pavement data set.A generation antagonistic network is designed and optimized by using coordinate index method.Finally,the network generates data images of asphalt pavement features which are replaced by soil pavement features.The mean square error,peak signal-tonoise ratio,structural similarity and similarity degree of the generated images are 3.563,28.520,0.753 and 0.292.The generated adversarial network solves the problem that the data sets cannot be transferred and applied due to the different pavement materials.Finally,in order to prove the validity of semantic segmentation network for pavement compaction scene segmentation after data set migration,further simulation tests are carried out.The reliability of lane keeping assistance system(LKA)has been proved by many scholars in experiments.In order to verify the error of road roller operating track controlled by road edge information recognized by semantic segmentation,the LKA system was improved according to the operating parameters of the roller.The road edges identified after semantic segmentation were imported into the simulation experiment platform.The maximum lateral errors of the final roller are 10 cm,13cm and 16 cm,18cm along two different tracks at the speed of 3km/h and5km/h.Finally,the experiment proves that the road edge information based on semantic segmentation is feasible to control the working path of the roller.