Automatic Camera Calibration And Automatic Establishment Of Frenet Coordinate System In Expressway Scenes

The expressway distributed video surveillance system is adopted to automatically obtain the spatio-temporal motion information of vehicle targets based on the road,and to construct a unified description of vehicle trajectories in a wide range of road areas,which has theoretical significance and practical application value for expressway traffic situation awareness and traffic operation optimization.The basic conditions for realizing the above tasks are automatic camera calibration and the establishment of the road time and space benchmark.However,the traditional automatic camera calibration method of traffic monitoring has great deficiencies in scene adaptability,calibration accuracy,and speed because of the various forms of expressways and the variable pose of many surveillance cameras,and can not be organically unified with the road.Accordingly,this work focuses on the automatic camera calibration and the automatic establishment of the Frenet coordinate system in expressway scenes based on the traffic surveillance camera calibration theory,and the following innovative achievements are achieved.(1)A method of road vanishing point estimation based on vehicle features is proposed.Aiming at the problem of expressway road vanishing point estimation,this paper designs a vehicle direction estimation network based on deep learning technology,and uses classification and the offset mechanism to transform the road vanishing point estimation problem into the vehicle direction estimation problem.On this basis,the accurate estimation of orthogonal vanishing points of a group of roads is realized by cascaded Hough transform.At the same time,a vehicle direction data set for camera calibration is constructed and released.The experimental results show that the proposed method can effectively improve the accuracy and stability of road vanishing point estimation.(2)A method of automatic camera calibration in expressway scenes is constructed.To solve the problem of low automatic camera calibration accuracy in different expressway scenes,the constraint models based on lane dotted line,lane width,and moving vehicles are established from the point of view of static and dynamic constraints.Then,combined with the road orthogonal vanishing point,a complete initial calibration condition is formed,and the three-dimensional reprojection model is used to iteratively optimize the camera parameters.The experimental results show that the proposed method has high calibration accuracy and is suitable for a wide range of expressway scenes.(3)A fast automatic camera calibration network based on scene learning is designed.To solve the problem of the complex and inefficient calibration process of traditional methods,a deep calibration network combining convolution attention mechanism is designed according to the mapping relationship between the global features of the scene and the camera parameters.The network integrates vanishing point estimation,feature extraction,and camera pose parameter estimation into the same process,and can realize fast camera parameter estimation directly through a single expressway image.At the same time,a multi-view camera calibration data set is constructed and released.The experimental results show that the deep calibration network has better calibration accuracy for ordinary straight roads and curved roads with small curvature,and the running time is much lower than that of the traditional methods.(4)The Frenet coordinate system based on the geometry of the expressway is established.To solve the problem of a unified description of vehicle trajectories in a large range of road areas,a coordinate system that accords with the shape of the expressway is established by introducing the concept of the Frenet coordinate system based on camera calibration,so as to accurately express the relative position relationship between vehicles and the road.For the automatic establishment of the Frenet coordinate system,this paper further proposes a method based on vehicle trajectories.Experiments show that the proposed method can establish a road-based coordinate system in different expressway scenes,and effectively describe the spatial coordinates and trajectories of vehicles relative to the road.The above researches focus on solving the bottleneck of expressway monitoring in automatic camera calibration and automatic establishment of the Frenet coordinate system,which realizes the organic combination of the existing calibration technology and the road.It lays a solid foundation for further establishing a unified road time and space benchmark under the condition of cross-camera and obtaining the motion parameters of whole vehicle targets.In addition,the two data sets effectively fill the deficiency of the lack of traffic surveillance camera calibration data and provide some data support for the application of depth networks in the field of traffic surveillance camera calibration.