Speed Measurement from Traffic Camera Video Using Structure-from-Motion-Based Calibration

Abstract

Speed measurement from traffic camera video is an active research problem, primarily due to the challenges associated with accurate camera calibration in an uncontrolled environment. Traditional calibration techniques often require prior knowledge of the scene or manual input from the user, limiting their applicability in real-world scenarios. In this work, we propose a novel approach for traffic speed measurements using a camera calibration based on automatic 3D scene reconstruction via structure-from-motion (SfM). Our approach leverages deep learning-based feature extraction and matching, specifically SuperPoint and SuperGlue, to achieve precise scene reconstruction. By placing the camera within the reconstructed environment, we obtain its intrinsic and extrinsic parameters without requiring predefined reference objects. This allows us to establish a reliable reference for measuring distances in the scene. With this setup, we can accurately measure point distances on the ground plane, enabling robust speed estimation of moving vehicles. We present the implementation of our speed measurement method as a realtime application based on YOLO11 object detector and BoTSORT object tracker. Our approach achieves speed measurement accuracy with an error of 6.9 km/h compared to a GPS RTKbased reference benchmark, demonstrating its effectiveness for traffic monitoring applications.