Graduation Project — Reality Capture from Multiple Video Cameras

This project addresses the knowledge gap in Intelligent Transportation Systems (ITS) by leveraging multiple video streams and advanced photogrammetry techniques to generate accurate 3D models of vehicles. By improving vehicle tracking and recognition, the project aims to contribute to more effective traffic management, emergency response systems, and road safety.

Despite challenges such as moving cameras, shifting lighting conditions, and processing requirements, the project emphasises optimising computational efficiency. YOLOv5 was used for object detection, while MIDAS was employed for depth estimation, both integrated through PyTorch. Feature detection was performed using the FAST algorithm, and 3D reconstruction was carried out with OpenCV.

Key methods included:

• YOLOv5 for precise real-time vehicle detection
• Camera calibration using chessboard patterns and MATLAB
• Depth estimation via MIDAS (DPT_Large)
• 3D reconstruction through aligned multi-angle point clouds

Although the point cloud was low-resolution, the project successfully demonstrated the feasibility of combining these methods to build partial 3D models of vehicles. These results serve as a foundation for further development in real-time ITS applications.

This project holds strong scientific, technological, and socioeconomic potential. It proposes a novel, efficient approach for 3D vehicle modelling that could be integrated into future ITS systems and commercial traffic monitoring tools.