Towards Large-Scale Video-Based Highway Traffic Monitoring

Traffic safety on highways is supported by a variety of technical measures, including countless camera systems that are often only monitored by human operators. However, due to the sheer amount of data, safety monitoring and accident prevention are limited by human resources. In this paper, we present an efficient system capable of extracting accurate vehicle trajectories from the vast amount of video data generated by modern highway infrastructures. Our proposed system conveniently leverages bird's eye view transformations estimated from aerial data or street marker geometry to generate geo-localized trajectories. Utilizing existing infrastructure, we demonstrate that the central data for video-based highway traffic monitoring can be reliably extracted. Remarkably, this can be achieved solely relying on uncalibrated cameras and noisy video streams.