Abstract Accurate camera calibration is essential for intelligent traffic monitoring, enabling precise vehicle tracking, speed estimation, and automated traffic flow analysis. However, traditional calibration methods - including manual, geometry-based, and deep learning approaches - face limitations in scalability, accuracy, and computational efficiency. This study proposes an improved framework for 3D scene reconstruction and traffic camera calibration based on a spherical camera model and an icosahedral image representation. In contrast to prior approaches that decompose panoramas into multiple perspective views, our method processes full 360° panoramic images using tangent image projections, which preserves geometric consistency and enables efficient feature extraction and matching. The framework is evaluated on the LUMPI dataset, focusing on focal length estimation accuracy, localization precision, and geometric consistency. Experimental results demonstrate that the proposed method achieves lower focal length estimation errors, improved localization accuracy, and consistent ground plane geometry. Furthermore, it significantly reduces computational costs, achieving a 94× reduction in feature matching time and a 26× speed-up in reconstruction. The proposed framework offers an accurate and computationally efficient solution for traffic camera calibration, suitable for large-scale, real-world applications.
Ruiz et al. (Thu,) studied this question.