Research On The Key Problems In Binocular Stereo Measurement System Based On Random Illumination

Three-dimensional (3D) shape measurement of real-world objects has been widespread applied inthe fields of aerospace, vehicle navigation, machinery manufacturing, biomedicine, heritagepreservation, gaming and entertainment, etc. With the modern industrial and scientific technologicaldevelopment, the requirements of3D point cloud measurement technology for measurement precision,measurement efficiency, data quality, measurement automation and stability are increasing. But howto improve the performance of3D point cloud measurement to meet the increasing requirements hasbecome a critical problem that needs to be addressed. The research in this thesis aims to improve themeasurement performance of binocular stereo measurement system based on random illumination.Several key issues in3D point cloud measurement based on images are further studied in this paper,including self-occlusion, local highlight effect, high-speed automatic3D reconstruction, automaticregistration of multiple range images, etc. The main contents and contributions are as follows.A stereo matching algorithm with high-efficiency and automation is put forward for the speckleimages. By combining the feature matching and area matching methods, an algorithm for automaticfeature extraction and stable feature matching is proposed to collect a number of pointcorrespondences. These correspondences are taken as the seeds of dense point matching propagatingprocedure, so it can not only solve the automatic implementation of the whole matching process, butalso effectively avoid point absence in discontinuous regions. Besides, the parallel accelerationtechnology based on GPU is adopted to rapidly obtain dense image point correspondences in thematching propagating procedure.The sub-pixel matching methods of dense image point correspondences are deeply studied. Twomatching algorithms are respectively proposed to improve the quality of the measurement data inedge and minutia regions. In the edge point matching algorithm, to alleviate the matching error causedby the intensity disturbance of background, a correlation matching algorithm with adaptive deformedwindow is put forward to improve the sub-pixel matching precision. The speckle area constraint isimposed to filter out false correspondences located in the background region, which improves therobustness of the algorithm and the quality of the reconstructed edge points. In the minutia matchingalgorithm, the classical correlation matching algorithm is first adopted to rapidly obtain3D pointclouds on the object's surface. Then according to the curvature properties of these reconstructedpoints, the improved correlation matching algorithm is applied in minutia regions again. Therefore, it improves the sub-pixel matching accuracy, and better reconstructs detail features on object's surface.To alleviate the effect of self-occlusion and local highlight of the random-illumination-basedbinocular stereo measurement system, a novel3D shape measurement method is proposed withmonocular/binocular measurement modes. In this method, binocular stereo structure is utilized toreconstruct the common field of view of the two cameras, and the unreconstructed area (data absence)is filled up by two monocular measurement units, which are composed of one single camera and therandom pattern projector. The binocular measurement unit and the two monocular measurement unitsworking together can make the measurement results more integrated and effectively avoid pointabsence caused by the highlight and occlusion in any single camera's view. A simple and effectivecalibration method is presented to locate the projected rays of the random pattern projection. It isunnecessary to determine the projection center or calibrate the distortion of projector lens, since thesystem precision is not affected by these issues. By searching and integrating the overlapped pointsbetween the three measurement datasets obtained in the binocular and monocular units respectively,the3D measurement data output in one measurement has no redundancy.A novel scheme for in-process registering multi-view scans with the aid of a miniature attitudesensor has been presented. Since the attitude sensor is fixed on the scanner during the measurementprocess, a simple yet effective algorithm is proposed for calibrating the relative attitudes (rotationtransformation) between the attitude sensor and the scanner. When the scanner is moved from onestandpoint to another in a measuring process, both the calibration results and the real-time readings ofthe attitude sensor are utilized to compute the rotation movement of the scanner. After applying therotation transformation to the current point dataset, the translation movement is efficiently determinedby exploiting the normal vector constraint between the correspondence points. Experimentsdemonstrate that the proposed registration can be done in-process with convenience and stability, andthe rigid transformation obtained can serve as a qualified initial estimate for further fine registration.By introducing multi-view geometry theory to binocular stereo measurement technology for thefirst time, a novel algorithm is proposed for automatically registering the measurement data takenfrom two orientations by a stereo sensor, without any instrument for assistance. The pointcorrespondences between two of the four natural intensity images, which are captured by the left andright cameras in the two orientations, are first obtained. Then the calibrated parameters of thebinocular stereo structure, together with the effective matching constraints between the four images,are utilized to calculate the relative pose of the stereo sensor based on the multi-view geometry theory.The proposed algorithm not only looses the registration condition, which is helpful to improve the measurement efficiency, but also improves the registration accuracy and robustness since itincorporates as much constraints as possible in the algorithm.The proposed algorithms have all been fully comparative analyzed and experimental verified, andmost of them have been applied in the ReCreator measurement system developed in our lab.