Camera Pose Estimation Based On Feature Tracking

Most automatic navigation techniques are based on simultaneous localization and mapping techniques.And the real-time and accuracy of camera pose estimation are very important in these techniques.Camera pose estimation is also widely used in mobile robot vision navigation and 3D reconstruction system.In the computer vision geometry,the mathematical model between 3D space camera and 2D images,includes the camera's intrinsic parameter matrix and extrinsic parameter matrix.The camera calibration can obtain the intrinsic parameter matrix of the camera,while the main purpose of camera pose estimation is to obtain the extrinsic parameter matrix of camera.Firstly,camera calibration is very important for camera pose estimation,because the accuracy of camera intrinsic parameter matrix will directly decide the speed and accuracy of camera pose estimation.In this thesis,we use a flexible and accurate camera calibration algorithm called Zhang Zhengyou's Camera Calibration Algorithm.In many camera pose estimation algorithms,camera pose estimation is usually achieved by matching the input image with the key frame.The camera pose estimation algorithm adopted in most simultaneous localization and mapping techniques are based on pixel tracking,which cannot be achieved in real time.Based on a large number of previous works by other researchers,this paper proposes the real-time camera pose estimation algorithm based on feature point tracking.The main innovations are as follows:1)We propose a hybrid approach based on Lucas Pyramid optical flow and FAST features.Optical flow method can accurately track the feature points in continuous images.FAST feature is a fast feature point extraction method.Compared with the ORB descriptor matching method,the proposed method not only overcomes the shortcoming that ORB matching cannot achieve real-time,but also has better accuracy than the matching effect.2)According to the method we proposes,we preprocess the interval of the gray values and positions of the feature points,and we deal with the gray interval and physical distance of feature points,and more accurate results are obtained.3)We set up the update condition of key frame by summarizing the experiment,and prove the effectiveness of the preprocessing by experiments.Finally,the camera pose estimation system is implemented in two cases of video and real-time camera.A large number of experiments show the algorithm designed in this paper can accurately and quickly estimate the camera pose of video and camera in real time.