Researches On Vision-based Pose Estimation And Point Target Motion Measurement

Pose estimation and point target motion measurement are two basic problems in vision-based measurement.They are widely used in aircraft navigation and landing,space rendezvous and docking,intelligent mobile robot navigation,augmented reality and ground target localization of UAV,flighting missile target localization,and so on.This dissertation studies on pose and motion parameters estimation and further improves speed,accuracy,robustness and adaptability of these algorithms.The main contributions of this dissertation are as follows:The orthogonal iteration algorithm for single camera pose estimation is explored in depth.The correct convergence proof of this algorithm is given for the first time.The accelerative algorithm and the angle error minimized orthogonal iteration algorithm are respectively proposed to increase the computational speed and to improve objective function.For the accelerative algorithm,the redundant computing is pointed out and regulated,so that,the complexity of each iteration is reduced from O(n)to O(1),and the computing speed is increased.For the angle error minimized algorithm,weighted absolute orientation is applied to update rotation matrix,and the precision is higher than the original algorithm when the reference point depths are in large range.The polynomial equations algebraic methods for pose estimation are studied,then,the direct least squares framework is proposed and applied to issues of pose estimation based on the straight line segment correspondences,multi-camera system pose estimation and hand-eye calibration.This framework can be used in the case where the objective function has the quadratic form of rotation.For the issue of pose estimation based on the straight line segment correspondences,a distance metric between two straight line segments is also proposed,which combines four factors including the distances between the end points,the distance between the middle points,the angle between the two lines and the lengths of the lines,These solutions are not only the closed-form and O(n)complexity,but also global optimal.The mirror-based camera pose estimation is studied for the case where the reference points are not directly in the field of view.The algorithm based on the reflection rotation averaging is proposed.The key idea is to interpret the reflection of mirror by rotation averaging,so that the rotation of camera can be precisely obtained.Hereby,the rotated rotation averaging is proposed,including the model of reflection rotation averaging.Then the two solutions of this model respectively based on SVD and quaternion are proposed to minimize the square chordal distance,and the sufficient and necessary conditions of the unique solution are proposed.These model and solutions extend rotation averaging theory.For monocular point target motion measurement,the partial observability is analysed in the case of static camera observing a moving target where the target motion is not fully determined.Then,a LSSVM regression based method is proposed.This method describes the target motion model by kernel function,and introduces the penalty factor to consider the model error.Compared to traditional monocular trajectory triangulation,it has more adaptability for different target motion models.Multi-camera trajectory triangulation is proposed to determine a point target trajectory in different conditions of time information.It aims at motion measurement in three conditions including known time,no time registration among multi-cameras and no time information.In contrast with traditional trajectory triangulation,it has more adaptability for different conditions of time observation.