In this paper, we present a reliable and scalable approach for real-time estimation of link speeds (i.e., traffic speeds on specific road segments) based on video feeds coming from monocular cameras. We detect and track vehicles of specific types, identify anchor points (or keypoints) on them, compute their poses, and use this information to estimate their speeds. We use deep learning methods for vehicle detection, tracking, keypoint detection and localization, and traditional 3D pose estimation techniques for which precise mathematical solutions are available. Thus, our approach exploits the best of both worlds. The proposed approach does not require any physical measurements (extrinsics) in the road scene, making it scalable and easy to install. Our results on video feeds from Bangalore, India, show that the method is able to generalize well for cameras mounted on street light poles, congested traffic situations, and various lighting conditions. Thus, the solution is suitable for emerging market scenarios where traffic tends to be chaotic and dense, and mounting speed sensors or strategically located downward-facing cameras is not feasible. The code and dataset for this work are being made available2.2https://github.com/ShantamShorewala/vehicle-speed-keypoint-data
Link Speed Estimation for Traffic Flow Modelling Based on Video Feeds from Monocular Cameras
2020-09-20
747907 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch