In recent years, with the development of autonomous driving technology, 3D recon- struction from street-view perspectives has become a focal point for many researchers. However, due to the characteristics of street-view perspectives, this task has consistently faced numerous challenges. Additionally, in practical scenarios, camera calibration accompanied by Global Navigation Satellite System (GNSS) / Inertial measurement unit (IMU) systems will incur higher equipment costs than camera-only systems. Mean- while, using only one monocular commercial camera for 3D reconstruction can also reduce the equipment cost of this task. On the other hand, the emergence of Neural Radiance Fields (NeRF) technology has introduced new approaches to the task of 3D reconstruction, leading to the proliferation of numerous algorithms based on NeRF of reconstructing 3D scenes. However, most existing research focuses on stan- dardized and calibrated datasets or is limited by the need for multi-view inputs or the inclusion of LiDAR data to deal with street-view reconstruction. In this work, a comprehensive workflow from street-view video to dense 3D re- construction with one monocular camera is developed. Using a commercial camera, this workflow aligns video capturing, camera calibration, pose estimation, and 3D reconstruction with evaluation for practical autonomous driving applications. To deal with the characteristics of street-view scenes, image segmentation and image inpainting are also involved in data preprocessing steps in this pipeline. Meanwhile, two different methods, Structure-from-Motion (SfM) and Direct Sparsity Odometry (DSO), are tested and compared for the camera calibration and pose estimation tasks. To address the problem of street-view reconstruction with calibration information, streetsurf is applied for our monocular dataset. The performance of such a GoPro dataset has achieved an average Peak Signal-to-Noise Ratio (PSNR) of 30.50, Structural Similarity Index (SSIM) of 0.934, and Root Mean Square Error(RMSE) of 1.30m. In summary, this thesis sets out a successful workflow for monocular vision-based 3D reconstruction in street-view scenes. This research provides valuable insights and directions for using uncalibrated commercial cameras in autonomous driving applications, addressing the practical challenges of real-world scenarios.     «

In recent years, with the development of autonomous driving technology, 3D recon- struction from street-view perspectives has become a focal point for many researchers. However, due to the characteristics of street-view perspectives, this task has consistently faced numerous challenges. Additionally, in practical scenarios, camera calibration accompanied by Global Navigation Satellite System (GNSS) / Inertial measurement unit (IMU) systems will incur higher equipment costs than camera-only...     »