Enhanced visual SLAM for surgical robots with cylindrical scene recognition in digestive endoscopic procedures

Abstract

Traditional SLAM (Simultaneous Localization and Mapping) methods struggle in natural orifice transluminal endoscopic surgery (NOTES) due to lumen diameter changes, which disrupt monocular feature tracking. To overcome this, we propose a SLAM algorithm with cylindrical scene recognition to enhance tracking in variable-geometry environments. Using cylindrical cycle optimization and principal component analysis (PCA), it estimates structures in real time, reducing spatial complexity and speeding convergence. Additionally, backward projection error correction improves landmark generation and scene continuity. Simulations show a 64 % gain in feature matching efficiency, and medical model tests confirm improved tracking accuracy, validating this approach for safer, more reliable NOTES navigation.