Master of Philosophy
School of Electrical, Computer and Telecommunications Engineering
Recently, three-dimensional (3D) shape measurement technologies have been extensively researched in the fields such as computer science and medical engineering. They have been applied in various industries and commercial uses, including robot navigation, reverser engineering and face and gesture recognition. Optical 3D shape measurement is one of the most popular methods, which can be divided into two categories: passive 3D shape reconstruction and active 3D shape imaging.
Passive 3D shape measurement techniques use cameras to capture the object with only ambient light. Stereo vision (SV) is one of the typical methods in passive 3D measurement approaches. This method uses two cameras to take photos of the scene from different viewpoints and extract the 3D information by establishing the correspondence between the photos captured. To translate the correspondence to the depth map, epipolar geometry is applied to determine the depth of each pixel. Active 3D shape imaging methods add diverse active light sources to project on the object and use the camera to capture the scene with pre-defined patterns on the object’s surface. The fringe projection profilometry (FPP) is a representative technique among active 3D reconstruction methods. It replaces one of the cameras in stereo vision with a projector, and projects the fringe patterns onto the object before the camera captures it. The depth map can be built via triangulations by analysing the phase difference between patterns distorted by the object’s surface and the original one.
Those two mainstream techniques work alone in different scenarios and have various advantages and disadvantages. Active stereo vision (ASV) has excellent dynamic performance, yet its accuracy and spatial resolution are limited. On the other hand, 3D shape measurement methods like FPP have higher accuracy and speed; however, their dynamic performance varies depending on the codification schemes chosen. This thesis presents the research on developing a fusion method that contains both passive and active 3D shape reconstruction algorithms in one system to combine their advantages and reduce the budget of building a high-precision 3D shape measurement system with good dynamic performance. Specifically, in the thesis, we propose a fusion method that combines the epipolar geometry in ASV and triangulations in the FPP system by a specially designed cost function. This way, the information obtained from each system alone is combined, leading to better accuracy.
Furthermore, the correlation of object surface is exploited with the autoregressive model to improve the precision of the fusion system. In addition, the expectation maximization framework is employed to address the issue of estimating variables with unknown parameters introduced by AR. Moreover, the fusion cost function derived before is embedded into the EM framework. Next, the message passing algorithm is applied to implement the EM efficiently on large image sizes. A factor graph is derived from fitting the EM approach. To implement belief propagation to solve the problem, it is divided into two sub-graphs: the E-Step factor graph and the M-Step factor graph. Based on two factor graphs, belief propagation is implemented on each of them to estimate the unknown parameters and EM messages. In the last iteration, the height of the object surface can be obtained with the forward and backward messages. Due to the consideration of the object’s surface correlation, the fusion system’s precision is further improved.
Simulation and experimental results are presented at last to examine the performance of the proposed system. It is found that the accuracy of the depth map of the fusion method is improved compared to fringe projection profilometry or stereo vision system alone. The limitations of the current study are discussed, and potential future work is presented.
Zhu, Yuewen, Three Dimensional Shape Reconstruction with Dual-camera Measurement Fusion, Master of Philosophy thesis, School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, 2022. https://ro.uow.edu.au/theses1/1713
FoR codes (2008)
0906 ELECTRICAL AND ELECTRONIC ENGINEERING, 1005 COMMUNICATIONS TECHNOLOGIES, 1006 COMPUTER HARDWARE
Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.