Research on 3D Reconstruction Method for Dynamic Scenes Based on Monocular Vision

This paper proposes a monocular vision-based 3D reconstruction method for dynamic scenes, which reconstructs 3D models of scenes from image data captured in moving environments. A dynamic scene refers to a situation where the speed of the camera carrier exceeds 10 km/h. Images captured in dynamic scenes often contain various types of noise, leading to significant errors in the 3D model, poor point cloud reconstruction, and excessive noise points. To improve reconstruction quality, this paper conducts an in-depth study of 3D reconstruction models and proposes an optical-inertial compensation-based 3D reconstruction scheme through information fusion. The method begins by detecting and matching features, eliminating mismatches, and selecting initial image pairs to initialize the model. Subsequently, the scene is reconstructed incrementally. During reconstruction, the system state is estimated by minimizing photometric errors, while the colors of map points are rendered. Finally, experimental results demonstrate that the proposed method achieves realistic 3D reconstruction with reduced noise points, the accuracy of the proposed algorithm in model construction, pose estimation, and model color rendering.