Semi-global matching measurement method based on single-frame continuous coded fringes

Objective Due to its low cost and operational simplicity, the binocular stereo vision system has become a widely adopted solution for three-dimensional perception in industrial automation. The point cloud generated through stereo matching provides real-time spatial pose feedback for production processes, whose quality directly depends on the completeness and accuracy of the point cloud. Researchers have thus focused on achieving complete and accurate stereo matching. Currently, Speckle Projection Profilometry (SPP) is a mainstream approach for this purpose. Unlike Fringe Projection Profilometry (FPP), which requires phase calculation from multiple coded fringe patterns for accurate matching, SPP offers advantages in both matching efficiency and projection cost due to its single-frame projection. However, speckle matching algorithms rely on the spatial uniqueness of speckle features, which imposes stricter requirements on projection quality. Moreover, the discontinuous nature of speckle coding often fails to provide sufficient feature information for accurate stereo matching at disparity step positions. To address these limitations, this paper proposes a semi-global matching (SGM) measurement method based on single-frame continuously coded fringes to meet the practical demands of industrial production.

Methods Based on the binocular stereo vision system, an independent fringe generator is added in this paper. A kind of single-frame continuous coded fringes related to SGM matching path is proposed (Fig.3). The SGM cost initialize is improved by increasing the investigation of the median and average gray values of the neighborhood pixels, in order to avoid the error caused by the interference of the central pixel (Fig.4). An adaptive sub-region algorithm is designed to realize two-stage parallel computing (Fig.1). The Gaussian pyramid is introduced into the core area for stratification (Fig.5), and disparity estimation is performed on each layer to reduce the traversed disparity range.

Results and Discussions In order to demonstrate the effectiveness and practicability of the proposed method, the experiments of feasibility verifying, three-dimensional reconstruction of weak texture regions, accuracy analysis at different distances and reconstruction of complex environments are carried out. At 1.0m, proposed method achieves an accurate matching with correct matching rate (ZNCC≥0.9) of 92.77% and RMSE of 0.89 mm, which is superior to ZNCC-based method and SIFT-based method (Tab.3). By comparing with two commercial depth cameras, proposed method behaves relatively accurate and stable plane sensing ability in a large depth range. The planar measurement error of the proposed method is 1.54 mm at 2 m and 3.89 mm at 4 m(Fig.11). The quantitative measurement of step samples shows that proposed method realizes depth measurement with MAE less than 1.25 mm and RMSE less than 1.45 mm at 2.0 m (Fig.12). This method also achieves high-quality matching and three-dimensional reconstruction of complex environments (Fig.13).

Conclusions Focusing on the issue that single-frame speckle structured light has high requirements on projection quality and its discrete coding characteristics are difficult to provide sufficient feature information at disparity step to complete accurate matching, a semi-global matching measurement method based on single-frame continuous coded fringes is proposed. By projecting single-frame continuous coded fringes associated with matching paths, sufficient feature information is supplemented for each path without increasing the computational complexity, which helps to achieve accurate stereo matching in weak texture regions. At the same time, cost initialize of SGM is improved based on the fringe to avoid the mismatch caused by the influence of projected fringes on gray level of center pixel. An adaptive partition and core region disparity prediction method is proposed. The Gaussian pyramid is introduced to estimate the hierarchical disparity. By reducing the disparity range of the traversal, the computational complexity is reduced. It achieves complete matching in complex environments and realizes precise measurement with the correct matching rate of 92.77% and the RMSE of 0.89 mm for a week-textured target at 1.0 m. For plane detection at different depths, the measurement error is 1.54 mm at 2.0 m and 3.89 mm at 4.0 m. Proposed method exhibits high precision and robustness.