A Real-Time Pipeline for Traffic Violation Detection and Analysis Using Monocular Video

Abstract

Reliable traffic violation detection is essential for improving road and pedestrian safety. Existing monocular vision-based methods often rely on pixel-level object representations, which limit spatial reasoning and real-world applicability. This paper presents an end-to-end, real-time traffic violation detection framework that operates in world coordinates, enabling spatially consistent violation analysis. The framework integrates transformer-based object detection with camera parameter estimation from image–world correspondences using a pinhole camera model, achieving a measurement error of approximately 20 cm and a speed estimation MAE of 0.467 m/s. The complete framework achieves 44 FPS, demonstrating practicality for real-time roadside monitoring.