Research On Automatic Camera Calibration Algorithm In Highway Scene

From the surveillance video of the highway,a lot of valuable traffic information can be obtained,which provides important decision-making basis for the early warning of traffic incidents and traffic design planning.The purpose of camera calibration technology is to establish the mapping relationship between images and threedimensional space.A stable technology of camera auto-calibration is an important guarantee for the extraction of rich traffic information.At present,the camera auto-calibration algorithms widely use the theory based on the estimation of the vanishing points in a pair of orthogonal directions in the traffic scenes,which becomes the researching focus.For the estimation of the vanishing point in the first direction,this paper uses the optical flow tracking algorithm to obtain the trajectories of the moving vehicles.Then trajectories are transformed into a diamond space using a Hough transform based on a parallel coordinate system and the position of the vanishing point can be estimated using cumulative voting.However,due to the roads bending,the estimation error of the vanishing point in the first direction is large.Besides,because of lacking necessary information,it is difficult to estimate the vanishing point in the second direction stably.In terms of the above problems,the works of this paper are mainly reflected as follows.Firstly,this paper uses a method of the vanishing point estimating based on Huff voting.Before the trajectories accumulate votes in the diamond space,they are screened to reduce the influence of the curved trajectories on the estimation of the vanishing point.Secondly,this paper proposes a framework based on a single vanishing point.In the traffic scenes,there are many problems,such as lacking horizontal direction markings and extracting horizontal lines difficultly because of low camera resolution.Also it is difficult for the camera to be stably estimated in the horizontal direction when the camera is in special postures.This paper uses the idea of enumeration and temptation.The camera is calibrated with the single vanishing point in the first direction.The method performs good validity,accuracy and stability in the laboratory scene.Based on the framework,different calibration algorithms can be used by selecting different calibration calibrators in the actual traffic scenes.Finally,this paper presents a camera auto-calibration algorithm based on the lane line model and vehicle wireframe model.For the method of using lane lines as calibrators,the difficulty lies in the stable extraction of lane lines.In this paper,the vehicle trajectories are combined with the image gradient to achieve a stable lane lines detecting method.For the method of using the vehicles as calibrators,this paper uses deep neural networks.It can be used for stable detection of vehicle targets in bad environments.Compared with the former,this method is more applicable.