Research And Validation On Vision Perception Algorithm Of Drivable Area On Structured Road

Under the background of increasingly severe road safety situation,intelligent vehicle related technologies are developing rapidly.Automatic driving technology can effectively improve road safety and greatly improve traffic efficiency.Intelligent driving system is mainly composed of three functional modules: perception,decision-making and control,environmental perception is the most critical module in intelligent driving system architecture,and is also an essential precondition for accurate decision-making and planning of automatic driving.The detection of drivable area is an important task in the module of environment perception,in view of the deficiency of the current research on the perception algorithm of drivable area,this paper extended the research on vehicle environment perception technology based on the sub-project of the national key research and development plan,and carried out the research on perception algorithm of drivable area based on vision on structured road.Based on the convolutional neural network,this paper proposes lightweight perception algorithms for dynamic driving surface and static lane lines to realize the detection of drivable area in structured road.Meanwhile,the effectiveness and feasibility of this algorithm are verified and analyzed through simulation and real scene data experiments.The main research work of this paper are as follows:Based on image semantic segmentation,a drivable road segmentation algorithm is developed for dynamic drivable road surface.Through the analysis of the detection task for drivable area,the algorithm was developed based on the Deeplabv3+ model architecture.The spatial pyramid pooling structure and deep separable convolution were used to improve the structure and lightweight optimization of the algorithm.Compared with the common semantic segmentation algorithms,the proposed algorithm achieves tens of times speed improvement with less than 5% precision loss.The MIo U index in the BDD100 K verification data set reaches 0.7917,and the reasoning speed in the low computing power GPU platform is as low as 6.7ms.Based on multi-task learning theory,a lane detection algorithm is developed for static lane lines on structured roads.According to the analysis of lane line detection requirements and the deficiencies of current lane line detection related algorithms,this paper realizes the development of lane line location detection and attribute recognition multi-task algorithm by sharing characteristic network parameters,so as to provide structured lane line information for intelligent driving,at the same time,in order to improve the accuracy of lane line fitting,RANSAC was used to optimize the lane perception results.Compared with the common lane detection algorithms,the algorithm in this paper provides the attribute information of corresponding lane lines on the basis of accurate lane positioning detection,so as to realize more fine-grained detection of the drivable area.Finally,design simulation and real scenario experiment to test and evaluate the algorithm.The self-vehicle perception plane is established.Based on the imaging principle of the camera,the representation of the perception information on the perception plane is realized through spatial coordinate transformation.In terms of simulation test,the perception algorithm developed in this paper is deployed in the Python/CUDA environment,the simulation environment is built in the simulation software Car Maker,and the joint simulation is realized through TCP/IP,and the drivable area perception algorithm in this paper is verified through a variety of road scenes.In the real scene test,the image data of the urban road driving scene was collected through the on-board camera,and the algorithm was verified in the offline environment.The experimental results show that the proposed algorithm can accurately perceive the drivable area on structured road.To sum up,this paper researches the vision perception algorithm of drivable area based on structured road,and proposes the drivable road segmentation algorithm and lane detection algorithm to realize the perception of the drivable area together,which provides a reference for the research of the visual perception technology of intelligent driving.