A handheld triocular multi-line laser 3D reconstruction method based on trifocal tensor point transfer and geometric constraints

• A TMLL 3D reconstruction method based on TFT and geometric constraints is proposed. • Coarse matching is achieved based on spatial consistency, and fine matching is carried out using TFT point transfer. • The average M-AcR of the LSCPs in the proposed TMLL method is 99.86 %, with an average M-ReR of 76.72 %. • The matching accuracy and robustness of LSCP in the multi-line laser system are improved. • Accurate matching of LSCPs is of great significance to the development of multi-line laser reconstruction technology. To address the issue of declining stereo matching accuracy and robustness of laser stripe centre points (LSCP) with increasing laser stripe (LaSt) lines in binocular multi-line laser (BMLL) three-dimensional (3D) reconstruction system, which leads to suboptimal reconstruction performance, a triocular multi-line laser (TMLL) 3D reconstruction method based on the trifocal tensor (TFT) point transfer and geometric constraints is proposed. First, a backlit azimuth circle calibration plate combined with the RANSAC 6-point algorithm is introduced to compute the TFT among the left, middle, and right views. Then, coarse matching of LSCPs across the three views is achieved based on spatial consistency, narrowing the search range for matching LSCPs in the middle and right images from the left image. Subsequently, fine matching of LSCPs is realised through TFT-based point transfer, obtaining the correct matching triplet for the LSCP in the left image from the middle and right images. Finally, global registration of local single-frame reconstructed point clouds is achieved using reflective circular markers, obtaining complete 3D data of the measured object. Compared to traditional BMLL reconstruction methods, the proposed method eliminates the need for laser plane calibration and instead achieves stereo matching through TFT point transfer and geometric constraints, avoiding the impact of laser plane calibration errors on stereo matching. Experimental results demonstrate that the proposed TFT calibration method achieves an average point transfer error of less than 0.74 pixel, exhibits superior matching performance across different LaSt distributions, scenes, and LaSt line counts, with an average matching accuracy rate (M−AcR) of 99.86 % and an average reconstruction rate (M−ReR) of 76.72 %. The reconstruction error for the standard block is less than 0.039 mm. The proposed method improves reconstruction quality and provides reliable technical support for applications in intelligent manufacturing, cultural heritage preservation and navigation and positioning. The method is implemented on a GPU using CUDA, and for a TMLL system with 15 LaSts, the average processing time per frame is 75.316 ms.