Traffic-Oriented Three-Dimensional Vehicle Reconstruction Using Fixed Roadside Monocular Camera Sensors

Area of Science:

• Computer Vision
• Intelligent Transportation Systems
• 3D Reconstruction

Background:

• Monocular cameras are cost-effective sensors for intelligent transportation systems.
• Extracting reliable 3D vehicle information from monocular cameras is challenging due to sensor limitations and moving vehicles.

Purpose of the Study:

• To present a traffic-oriented 3D vehicle reconstruction framework using monocular image sequences.
• To improve the accuracy and efficiency of 3D vehicle reconstruction from roadside camera data.

Main Methods:

• Jointly exploiting semantic and non-semantic vehicle feature points for structural consistency and surface completeness.
• Utilizing a feature-map-consistency-based optimization strategy to refine feature point localization and reduce reprojection errors.
• Developing an optimized incremental Structure-from-Motion (SfM) pipeline with traffic-aware initialization, keyframe selection, and local bundle adjustment.

Main Results:

• Reduced mean reprojection error by 13.6%.
• Shortened reconstruction time by 43.9% compared to existing incremental SfM systems.
• Demonstrated effectiveness on real-world traffic surveillance videos.

Conclusions:

• The proposed framework offers a significant improvement in 3D vehicle reconstruction accuracy and efficiency.
• The method effectively addresses challenges associated with monocular camera data in intelligent transportation systems.
• This work contributes to more reliable 3D traffic data acquisition for intelligent transportation applications.