Research On Drivable Area Detection Algorithm Of Based On Multitasking Network And Network Distillation

With the development of transportation,the situation of road traffic accidents has become more and more severe,causing huge losses to the safety of people’s lives and property.Intelligent Connected Vehicle can take on part or all of the driving tasks and improve driving safety.Among them,the driving area identification technology can detect the current road area where the vehicle can drive to ensure safe driving on the road,which is the basic research work of Intelligent Connected Vehicle environment perception.Therefore,research on the identification technology of the driving area of Intelligent Connected Vehicle is of great significance for improving traffic safety.With the rapid development of artificial intelligence technology,Intelligent Connected Vehicle have become one of the research hot spots.The intelligence and connectivity of Intelligent Connected Vehicle are closely related to the vehicle-road coordination system.For Intelligent Connected Vehicle in a vehicle-road collaborative environment,achieving "full perception" of the road environment is crucial.Among them,the drivable area recognition technology,as the core component of the environment perception technology,has achieved some research results,but there are still improvements,specifically including:1)The application scenarios of the drivable area detection method based on lane lines and road points are limited;2)It is difficult to balance the accuracy and real-time performance of the driving area identification method based on deep learning;3)There is less research on the identification technology of the drivable area of the intelligent roadside system(road end);4)Existing drivable area identification algorithms are difficult to deploy in practical applications.Based on the above analysis,this article is based on computer vision and deep learning and other related technologies to carry out research on the identification algorithm of Intelligent Connected Vehicle based on multi-task network and distillation.The specific research content of the thesis is as follows:(1)Vehicle-side driving area detection based on multi-task network.This article deeply explores the application of semantic segmentation technology in driving area detection.Based on the semantic segmentation model Deeplabv3+ and BiSeNet,a multi-task network of road segmentation + scene recognition is constructed.First,from the environment in front of the current vehicle,the multi-task network extracts the drivable area suitable for the self-driving vehicle.Secondly,a post-processing algorithm of DBSCAN density clustering is proposed,which divides the drivable area into three parts: the main drivable area,the left drivable area and the right drivable area,so as to realize the drivable area identification function of the Intelligent Connected Vehicle.On the BDD100 K data set,the BiSeNet multi-task network proposed in this paper has an average MIoU(Mean Intersection over Union)of 79.37% and an FPS of 131.(2)Based on the optimization of the driving area detection model of the network distillation vehicle.It is difficult for deep learning networks to balance accuracy and real-time performance.To this end,this article adopts the knowledge distillation training method,uses Deeplabv3+ with a complex model structure and high accuracy as the teacher network,and uses BiSeNet with a simple model structure and good real-time performance as the student network to improve the accuracy of the student network through the teacher network.In addition,this article uses the self-attention distillation training method to further improve the accuracy of the model.On the BDD100 K data set,the knowledge distillation + self-attention distillation method proposed in this paper has an average intersection ratio of 84.05% than MIoU,and the identification accuracy of the drivable area is about 5% higher than that before optimization.(3)Intelligent roadside system(roadside)can identify the driving area.Only relying on the vehicle end to detect the drivable area,there may be blind areas of perception,such as the front vehicle or building obscuration,sensor detection range limitation,etc.;The intelligent roadside system(roadside)can perceive the identification blind area of the car side to realize the fusion perception of the road scene.Therefore,on the basis of the multi-task network model of the driving area at the vehicle end,this paper retains the feature sharing network and road segmentation branch,simplifies the scene recognition branch,and constructs the road end identification data set of the drivable area,and builds the road end driving area identification model.(4)Acceleration and transplantation of the driving area identification model based on the embedded platform.On the basis of the vehicle-end-road-end drivable area identification model,the acceleration and transplantation methods of the drivable area identification model are studied,and the vehicle-end-road-end drivable area identification model in this paper is deployed to the Jetson Xavier NX embedded platform.It is verified by experiments that the driving area identification algorithm proposed in this paper can run well on the Jetson Xavier NX embedded platform,with a good balance of accuracy,real-time and robustness.