Research On Camera Self-Calibration Based On Moving Vehicles

With the development of urban transportation,the accuracy of camera calibration becomes more and more important for traffic surveillance video analysis.However,the existing camera calibration methods have several limitations.First of all,the traditional method requires a specific calibration plane,which is not suitable for traffic monitoring scenarios.As for self-calibration method,it needs to rely on Manhattan world assumption,traffic zebra crossing or pedestrian presence.However,such information is not always available in traffic scene.Considering the problems of existing methods,in this thesis,a novel and practical self-calibration system for traffic surveillance camera is proposed to overcome the limitations of existing methods.Firstly,a camera parameter calculation method based on vehicle geometry feature is proposed.Background subtraction with Otus thresholding method is adopted,combined with the morphological operations and shadow removal,which object detection result can be refined.Based on the object foreground mask,Canny operator is introduced to perform edge detection and the vehicle geometric features are obtained by Hough transform.Next,mean shift clustering and Laplace linear regression are used to improve the accuracy of vanishing points.Furthermore,a method for calculating camera parameters based on vanishing points is proposed.The experimental results on different datasets show that the proposed camera parameter calculation method is effective and robust.Secondly,an anti-interference schema for camera self-calibration system based on 3-D vehicle model is proposed.The effective projection features of the 3-D vehicle model in the image plane are extracted to reduce the noise impact when occlusion occurs.Meanwhile,the estimation of distribution algorithm(EDA)is introduced to solve the constraint problem of unknown camera intrinsic parameters and the iterative evolutionary computation is used to search the local optimal solution for parameter optimization,thus reducing the re-projection error and improving the anti-interference capability of the self-calibration system.Experiments on different scenarios fully prove that the proposed anti-interference algorithm can improve system stability and anti-interference ability and the proposed system can finish the camera self-calibration by only relying on the moving vehicle in the traffic video.