Validating AI-Based Depth Estimation for Road Scene Reconstruction Using Dashcam Images and Low-Cost LiDAR
Bilguunmaa Myagmardulam, Kazuyoshi Takahashi

Toyama Prefectural University, Toyama, Japan.
Nagaoka University of Technology, Niigata, Japan.

Abstract

Monocular depth estimation powered by deep learning has the potential to significantly reduce the cost and complexity of 3D scene reconstruction in mobile applications. In this study, we conduct a foundational evaluation of the VGGT model^s performance using low-cost Mobile Mapping System (MMS) data, focusing on snow-free urban road environments. The MMS setup consists of a dashcam, GNSS/IMU, and 3D LiDAR sensor, enabling both image capture and 3D point cloud generation at low cost.
Lens distortion in dashcam images was corrected using MATLAB-calculated camera parameters before inputting the data into the VGGT depth estimation model. The resulting 3D structures were compared with LiDAR-based point clouds in CloudCompare to assess spatial accuracy. Structure-from-Motion (SfM) outputs from VGGT model were also used for cross-validation.
Our results show that, even under non-snowy conditions, monocular depth estimation can generate meaningful 3D structures for road scenes. These findings support the broader application of such models to regular road monitoring and infrastructure inspection. With further refinement, especially regarding accuracy thresholds acceptable to field operators, this approach could enable low-cost road patrol systems where only critical areas are followed up with high-end equipment. Future work will extend this validation to snow-covered scenes.

Keywords: Monocular Depth Estimation, VGGT, Low-Cost MMS, LiDAR, SfM, Dashcam, Road Scene Reconstruction, Snow Monitoring

Topic: Topic B: Applications of Remote Sensing