DOE-based structured-light method for accurate 3D sensing

This paper presents a compact and accurate three-dimensional (3D) sensing system that employs a diffraction optical element as a projection device. Compared with the conventional laser speckle-based 3D sensing methods, a gridline pattern is utilized instead of a dot pattern. The proposed pattern is designed according to a pseudorandom coding scheme, and eight geometrical elements are embedded into the grid cells to form a unique codeword for each defined grid-point. By extracting the grid-points with the proposed feature detector, a topological graph is established to separate each pattern element. A convolutional neural network is trained for robust identification of the projected pattern elements. Finally, a codeword-correction procedure is applied to refine the decoding results. Using the proposed system-calibration method, accurate 3D reconstruction can be realized for the decoded grid-points. The measurement of the planarity and step distance has an absolute mean error of only 0.2-0.3 mm, indicating that it is far more accurate than the measurement using classical laser speckle-based 3D sensors. To demonstrate robustness of the proposed decoding algorithms, targets with plentiful color and texture are used. The results show that most of the grid-points can be robustly detected and that complex surfaces such as human faces and bodies can be precisely reconstructed.