3D Scene Reconstruction for Robotic Vision Systems: An Approach to Ancient Artifact Reconstruction Based on 3D Gaussian Splatting

To address the high cost and low efficiency of existing ancient artifact reconstruction methods, the study presents a 3D scene reconstruction framework based on 3D Gaussian Splatting (3DGS) to support robotic vision and 3D Reconstruction in smart surveillance systems. Ancient artifact videos are first captured using a smartphone or standard camera. Utilize FFmpeg to rapidly extract video frames (extracting two frames per second), then leverage COLMAP’s capabilities of feature matching, structure from motion (SfM), and keyframe selection to directly feed the video frame sequence as input and output a multi-view image set with camera poses. A sparse point cloud is subsequently generated using the SfM algorithm and employed for scene optimization. Experiments on an ancient artifact dataset demonstrate the superiority of the proposed method over existing approaches. Compared to neural radiance fields (NeRF), the 3DGS model achieves substantial performance gains. Specifically, it improves the average peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) by 65.64% and 68.09%, respectively. The learned perceptual image patch similarity (LPIPS) score is reduced by 95.26%. Moreover, the average reconstruction time is shortened by 98.04%. The resulting 3D models exhibit high-fidelity texture restoration and accurate geometric representation. These findings provide a technical reference for achieving low-cost, high-quality 3D reconstruction of ancient artifacts.