Geolocating Traffic Signs using Crowd-Sourced Imagery

Abstract

Action cameras and smartphones have made it simple and cheap to collect large imagery datasets from the road network while driving. At the same time, several frameworks, e.g., Detectron2 and the TensorFlow Object Detection API, have made it fairly easy to build object-detection models for your imagery datasets. In this paper, we use the Detectron2 framework to detect 18 different common traffic signs from 351.469 images. The purpose is to automate the asset management of traffic signs in large road networks. A task that today often is done in a manual and labor-intensive manner. To improve the accuracy of determining the locations of traffic signs, we develop a new, general method that uses the size of the object detected (in pixels) and the camera's GPS position and heading. To further enhance the accuracy, multiple detections of the same physical traffic sign are clustered. The traffic-sign type and computed location are stored in a spatial data warehouse. The clustered locations are presented on a digital road network in a web app. This app allows visual inspection of the overall approach. We demonstrate that the accuracy of the computed locations is good, e.g., signs are placed on the correct side of the road or in/out of a roundabout.