Research On Measurement Method Of Rigid Object Position And Orientation Based On Monocular Vision

The pose estimation of moving targets is one of the hotspots in image processing,pattern recognition,photogrammetry and related subjects.With the urgent demand on monitoring the position and orientation of the moving targets,many scholars have developed various hardware systems and software platforms according to the different requirements.The measurements of position and orientation for monocular vision is widely concerned due to its non-contact with moving objects,simple structure,convenient and flexible operation and so on.According to the requirement of pratical application,the research on position and orientation measurement of rigid moving target for monocular vision is carried out in this thesis,including the position and attitude measurement of the moving target based on the cooperative and non-cooperative targets.The main contents of this thesis includes four folds:(1)In view of the actual application requirements,the research of the visual position attitude measurement for the cooperative target of monocular vision is carried out.To be specified,a camera on the ceiling is used to vertically shoot the moving targets on the ground.It needs to obtain three parameters of the position and orientation of moving target on the plane in real time,and the moving target has the premise of setting the coded marker.According to the top characteristics of moving targets,we design two kinds of codes,namely,circular coded marker and combined coded marker.The camera takes the image of the cooperative target with coded marker,and then extracts the coded marker in the image using the structural features of circular coding and the combination coding.As a result,the initial position and attitude of the moving targets is estimated from these coded markers.We use the initial position and attitude of the standard contour template and the moving object contour in the image to carry out two-dimensional ICP transformation to solve the precise position and orientation of the moving targets.(2)In some scenarios,moving objects do not have the conditions to construct cooperative coding marks.We conduct the research on the position and orientation measurement of non-cooperative moving targets to solve this problem.Because of the lack of easily recognizable features such as constructing coded markers,the 3D geometric model and camera parameters of target are required as prior information in the non-cooperative moving targets' measurement of position and orientation.With these prior information,OpenGL is used to generate the simulated images of moving targets in different position and orientation.The relationship between simulated images and real images is adopted to solve the position and orientation of moving objects.In this situation,the estimation of moving object position and attitude includes two parts:initial pose acquisition and accurate pose calculation.In this thesis,a method of coincident ratio of target region is proposed to obtain the initial value of position and orientation of non-cooperative moving targets.The target region feature set is constructed with the off-line position and attitude of moving target,and the initial position and posture are obtained by the relationship between the feature set of the moving target region and the target region feature of real image.The accurate pose calculation of non-corporative moving targets is characterized by the distance between the initial object contour of the OpenGL simulated image and the real image.By means of constructing the cost function of the contour distance,the precise position and attitude of moving target is solved iteratively through the nonlinear optimization.(3)The extraction of target contour is one of the most important steps to solve the position and orientation of the moving target.Exploring the efficient method for contour extraction has great influence on the reliability of pose estimation.The method based on active contour model has the advantage of automatically processing topological changes,capturing local deformation and extracting the complete target contour.Therefore,an adaptive active contour model based on region is proposed to extract the contour of moving target.The model is composed of global terms and local terms.The global item guarantees the segmentation efficiency and anti-noise ability of the model,and combines the local entropy information to construct the local energy term,and the evolution of curve is controlled in the uneven gray area.At the same time,according to the characteristics of local entropy reflecting the change of gray level in the neighborhood,the local entropy information of the image is used to adaptively determine the parameters of local and global terms.The global term plays a leading role in controlling the evolution of the curve far away from the edge position,and the local term leading curve evolves near the target contour,and the target contour is extracted by the synergistic of global term and local term.(4)The initial pose measurement of non-cooperative targets is determined by the overlapping ratio of target area,and its reliability is restricted by the characteristics of moving target region.In this thesis,we propose a target region extraction method based on Gaussian Mixture Model(GMM).The GMM superpose the Gaussian distributions by the mixing coefficients,assuming that the image contains the object and background with Gaussian probability distribution by unknown parameters.The expectation maximization(EM)algorithm is used to estimate the parameters and the Bayesian rule is utilized to extract the target region.Unlike the traditional Gaussian mixture model,the neighborhood information of the image is taken into account when the posterior probability is calculated to determine the target and background,and it has good robustness for the extraction of the target region of the intensity inhomogeneous images and noisy images.