Research On Identification And Location Of Object Based On Monocular Vision

Vision technology has been widely used in industry, medicine, aeronautics and astronautics,agriculture and other fields, with its non-contact, low-cost, and easy to achieve. It has been moreresearch and application, along with the rise of the computer and image processing hardwarecost-effective. The dissertation entirely researched the camera calibration, image feature matching,depth estimation and pose parameters, all of which were the key problems of the monocular visionsystem. The theories were used for pose parameters determination on the process of spacecrafts’rendezvous and docking and industrial online identification system.Image feature extraction and image matching are the basis of depth estimation,3D reconstructionand so on. A fast contour feature extraction method that is suited for the industrial online recognitionis proposed. About the image matching, the SIFT and other two local feature matching methods areresearched, combined with the shape context descriptors, a improved shape context descriptor isproposed and used for image matching. The algorithm is particularly applicable to the mismatchingprocess because of the similar or symmetrical target.The two-step calibration of Tsai and Zhang method based on the moved plane template areanalyzed, and an algorithm of linear calibration based on cross ratio invariance is proposed. Themethod avoids the nonlinear optimization of above calibrations, and only need adopt a linearcalibration, so that can be used for fast calibration of online detection system. On the basis of theabove camera calibration and improved shape context, a depth estimation method based on a bifocalmonocular vision is proposed. It solves the target depth estimation problem under the condition ofmonocular vision, the experimental results show that can satisfy the required precision, and themethod is particularly suitable for solving the smaller depth estimation issue.Monocular vision technology can be used for the location of the space object, that is, todetermine the relative pose relationship between camera coordinate system and the object coordinatesystem. The paper intensive studies the location issue of artificial flag image and no artificial flagimage under the monocular vision. First, object location algorithm of single image, two images andmultiple images with artificial flag are analyzed. Then the object location algorithm of single imagewith no artificial flag is researched implemented, and two algorithms are proposed. The first is thetarget location based on the known long one side of parallelogram according to the geometricrelationship. The second is the target location based on the Conformal Geometric Algebra, the poseparameters are obtained based on the objective function of feature points and feature lines that shouldbe optimized. The experiments on the object of industrial online identification system show that theproposed single image positioning has high precision and good real-time.The relative attitude parameters are firstly needed to determine when spacecrafts rendezvous anddocking. Three relative position and orientation estimation methods are proposed based on monocularvision. One is based on the depth estimation in monocular vision. By using three non-colinearreference points and monocular vision, the computation method for the relative position and attitudeparameters between spacecrafts is proposed. Firstly, the relative position parameters were computed using the bifocal imaging algorithm, then the relative attitude parameter were calculated by theposition parameters which has been obtained, at last the analytical solution of relative position andattitude parameters was achieved. Another method is used of quaternion based on monocular vision.The five feature optical points are made into two groups, three symmetry structure points as one group.The two equations sets of relative pose parameters are obtained by quaternion, then a closed-formsolution is derived, and correct terms are added to obtain more accurate estimates for the poseparameters. The third method is the use of monocular vision and laser range finder hybrid systeminstead of single monocular vision. The6D pose data measured by monocular vision and the highprecision ranging data from the laser range finder are fused. Compared with above two methods thatonly using single monocular vision to estimate, the method can further improve the estimationaccuracy of the position and orientation parameters between two spacecrafts.Monocular vision technology can be applied to the industrial online identification system. Thedissertation describes the whole recognition system that includes the hardware construction andsoftware implementation. A key step in contour extraction of cigarette image processing, a novel“two-step” method is proposed to copy with the different number of cigarette images in the picture.First step, global threshold segmentation and Blob analysis are used to distinguish the cigarette imageand the belt background, and then the smallest rectangle method is used to extract the cigaretterectangle contour. The above step extracts the rough contours of “single cigarette” and “multiplecigarettes”. Second step, global threshold segmentation, Blob analysis, and morphology are combinedto amend the contour extraction, to ensure the accuracy.The feature extraction and image recognition are needed after the cigarette contour extraction.On the basis of accurate cigarette extraction, the color, texture and shape image features in thecigarette contour should be extracted. Compared with the corresponding features in image featuredatabase that has been established, then the matching similarity is calculated to complete featurematching. Good running condition in tobacco logistics center that verifies the stability and practicalityof cigarette online identification system.On the basis of the system, a camera using for cigarette online identification is calibrated basedon the calibration algorithm in chapter3. Then spatial orientation is completed based on no artificialsigns single image location algorithm. On the one hand, the correctness of algorithm that proposed inchapter4is verified. On the other hand, effective methods are provided for target location forintelligent sorting system that using by single image.