Benchmark Dataset and Deep Model for Monocular Camera Calibration from Single Highway Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Surveillance Systems

Background:

• Single-image camera auto-calibration is crucial for traffic perception efficiency.
• Existing methods struggle with limited real-world data and multi-view scenario adaptability.

Purpose of the Study:

• To present a systematic solution framework for single-image camera auto-calibration.
• To address the challenges of data scarcity and poor adaptability in multi-view traffic surveillance.

Main Methods:

• Constructed a large-scale synthetic dataset (336,000 frames) using CARLA 0.9.15 with diverse highway scenarios.
• Developed DeepCalib, a deep calibration network utilizing triplet attention for vanishing point localization and camera pose estimation.
• Implemented a progressive learning paradigm: pre-training on synthetic data followed by fine-tuning on real-world data.

Main Results:

• DeepCalib achieved an average calibration precision of 89.6%.
• The method demonstrated a processing speed of 10 frames per second.
• Showcased robust adaptability to dynamic calibration tasks across various surveillance views.

Conclusions:

• The proposed framework effectively enhances camera auto-calibration in traffic surveillance.
• DeepCalib offers a significant improvement over conventional multi-stage algorithms in terms of speed and adaptability.
• The synthetic dataset and progressive learning approach contribute to better real-world performance.