ML-Based Road Asset Geolocation Using Object Detection and Camera Displacement

Geolocation methods identify objects in images and determine their geospatial locations. Current object geolocalization methods face challenges such as high hardware costs, limited object class coverage, difficulties with repeated object occurrences, and performance issues in dynamic environments. This paper introduces a machine learning approach for geolocalizing objects from low frame rate video using a single camera and image metadata, aiming to reduce costs and complexity compared to traditional methods. The method combines camera displacement data and object bounding boxes obtained from an object detection model to estimate geospatial locations. The approach was evaluated using diverse datasets that capture various driving environments and object types, demonstrating its capability to handle multiple scenarios.