Research Of Targets Positiong System Based On Monocular Camera Calibration

Location technology is a key technology in transportation,self-driving,IOT(The Internet of Things),etc.However,in some scenarios,GPS and other satellite positioning technologies are of poor effect and need the assistance of other positioning technologies.With the rapid development of image processing technology,positioning technology based on computer vision has received more and more attention due to its advantages such as high stability,low cost and wide using range compared with other traditional positioning technologies,and at present,it has become a hot research direction of positioning technology.Camera calibration is a key technique in visual positioning.However,the existing camera calibration methods have some limitations:the traditional camera calibration methods rely on the calibration plate,and the self-calibration methods need specific markers or require targets to travel along a straight line.Aiming at these problems,we improve the existing pedestrian tracking algorithm,and propose two camera self-calibration methods based on vehicle and pedestrian.The main research contents and innovative achievements of this subject are as follows:(1)A vanishing point estimation method based on vehicle 3d modeling and key points detection is proposed.First,vehicle targets in the picture are detected by Deep MANTA network and the 3D vehicle modeling is carried out with the vehicle 3D template database.Then,the vanishing points are estimated based on the 3D vehicle model,and the outliers and noises are processed by using the RANSAC(Random Sampling Consistent)algorithm and the mean shift algorithm to improve the accuracy of the vanishing points estimation,so as to complete the camera calibration.(2)A pedestrian tracking method combining appearance feature and movement feature is proposed.And based on this tracking method,a camera calibration technology is proposed.First,the YOLO detector is used to detect the pedestrian targets from the picture and extract the depth feature of the targets.Next,the movement feature extracted by Kalman filter and the appearance feature with color and depth feature are combined to match the same target in the adjacent frames.The vanishing points are estimated according to the condition that the height of one pedestrian is constant,and the mean shift algorithm and the RANSAC algorithm are used to deal with the noise and outliers,so as to complete the camera calibration.(3)A camera parameters optimization method based on minimum projection error is proposed.In order to remove the limitation of parameters in the process of camera calibration,we expand the range of parameters and use the EDA(Estimation of Distribution Algorithm)to screen out the parameter combinations with the minimum reprojection error,so as to improve the accuracy of camera calibration.The test results on different databases prove that the two camera calibration methods proposed in this subject have higher accuracy,robustness and universality than the existing automatic camera calibration methods.