In response to the challenges of insufficient texture modeling ability and limited real-time performance in traditional SLAM (Simultaneous Localization and Mapping) technology for industrial digital twin scenarios, this paper proposes a vision-based SLAM algorithm that integrates neural radiance fields and dense geometric priors. The method is based on the ORB-SLAM3 (Oriented FAST and Rotated BRIEF SLAM 3) framework, achieving high-fidelity scene reconstruction through differentiable rendering under dynamic voxel grid constraints. By combining multi-resolution hash encoding and CUDA (Compute Unified Device Architecture) parallel acceleration strategies, the modeling efficiency is significantly improved. Experiments show that on the TUM (Technical University of Munich) dataset, the algorithm achieves an average frame rate of 18.6 FPS (Frames Per Second), over 300 times faster than iMap (Implicit Mapping) and NICE-SLAM (Neural Implicit Compressive Encoding SLAM). The absolute trajectory error (ATE) RMSE remains below 0.0521 meters, approaching traditional SLAM performance levels, while simultaneously filling point cloud holes and reducing mapping time by 37.6%. Further validation reveals that the algorithm enables synchronized output of virtual environment models with complete geometric details and accurate physical attributes through co-optimization of dense point clouds and neural radiance fields, providing critical technical support for smart factory digital twin systems. The real-time interactive performance exceeding 18 FPS supports dynamic production line responses and human-machine collaboration optimization, while the dynamic voxel grid architecture and parallel computing design establish foundations for real-time digital twin modeling in large-scale industrial scenarios. This research offers an innovative solution for high-precision, high-efficiency virtual-physical mapping requirements in intelligent manufacturing.
Wang et al. (Thu,) studied this question.