Multi-Modal Lightweight 3D Transformer for Target Recognition and Reconstruction in Intelligent Construction

Abstract

In response to the dynamic occlusion, lighting changes, and computational efficiency faced by 3D object recognition in the intelligent transformation of construction engineering, this study proposes an optimized algorithm that integrates the Improved Outdoor Dynamic Scene Graph (ODSG) with a Lightweight Transformer. The SFM trajectory decoupling algorithm enhances geometric constraints and utilizes the S2DNet network to extract deep features, thereby optimizing the 3D reconstruction process. Additionally, a three-stage Lightweight Transformer model is developed, integrating self-supervised depth estimation, feature selection, and multimodal fusion mechanisms. The research results showed that on the threedimensional benchmark dataset, the optimized outdoor dynamic scene image framework achieved a dense reconstruction accuracy of 32.29% at 1cm precision, which was 9.53% higher than that of the traditional Collection of COLMAP system. The F1 scores reached 4.23 and 54.34 at 1cm and 5cm precision, respectively. In terms of object recognition, the optimized 3D Transformer achieved an average accuracy index of 25.33%, 17.68%, and 14.72% for 3D object detection on joint datasets in Easy, Mod, and Hard modes, respectively, which was 2.38% higher than that of the Monocular 3D Object Detection with Flexible Representations. The average precision for bird's eye view reached 36.15% in Easy mode, representing a 36.1% improvement over the conventional M3D-RPN baseline (26.56%). The research provides an efficient 3D perception solution for monitoring and safety warning of automated equipment in intelligent construction.