High-accuracy multi-camera reconstruction enhanced by adaptive point cloud correction algorithm.

• Multi-vision system is powerful, but unavoidable errors emerge in global calibration. • Global calibration error can be corrected by registration of the common parts. • The features of the targets are more adaptive than those of the standard objects. • The corrected models reach higher accuracy than that without correction. • The mean absolute error is 1.06 mm, and the mean relative error is 0.52% respectively. Multi-camera schemes can effectively increase the perception range of vision systems compared to single-camera schemes and are common in many optical applications. Unavoidable errors emerge in the global multi-camera calibration process, however, such as manufacturing error of the optical devices and computational error from marker detection algorithms, which drive down the accuracy of the camera system correlation. This paper discusses the causes of global calibration errors in detail. A four-camera vision system was built to obtain the visual information of targets including static objects and a dynamic concrete-filled steel tubular (CFST) specimen. Local calibration and global calibration were applied successively to realize multi-camera correlation, followed by filtering and stitching operations to acquire filtered global point clouds. A point cloud correction algorithm is designed accordingly to optimize the stitched point cloud structures and further improve the accuracy of the reconstructed surfaces. Based on the density features of the targets themselves (rather than standard calibration markers), the proposed point cloud correction algorithm is effective for various targets and adaptive under dynamic conditions. The point clouds and corresponding reconstructed models are shown to be more accurate after the proposed enhancement process. The point cloud correction algorithm also has strong adaptability to different static targets with complex surfaces and performs well under uncertain geometric changes and vibration. The results presented here provide both theoretical and practical support for advancements in multi-vision applications such as optical measurement, real-time target tracking, quality monitoring, and surface data acquisition. [ABSTRACT FROM AUTHOR]