Automated geolocalization of vehicles from UAV footage: evaluating measurement precision of object detection and segmentation methods

Abstract

Modern Road Traffic Monitoring (RTM) systems rely on advanced and precise technologies. Unmanned Aerial Vehicles (UAVs) coupled with state-of-the-art computer vision methods offer great utility in intelligent traffic monitoring and road safety analysis. However, the precision of these cutting-edge technologies is still under debate due to technical complexities, such as inaccurate road-user localization resulting in overestimated bounding dimensions, which could hinder their effectiveness in real-world scenarios. This research introduces a geolocalization method combining a feature-matching algorithm, SIFT, for automatic georeferencing of UAV frames with deep learning-based object detection and segmentation models. The study focuses on finding the most precise solution for vehicle geolocalization, preserving the vehicle shape and dimensions. The study explores three different configurations of YOLO object detectors: a standard YOLOv8 model, a hybrid model that integrates YOLOv8 with the Segment Anything Model (SAM), and a YOLOv8 variant that employs Oriented Bounding Boxes (OBB). The evaluation of results is focused on the dimensional accuracy, internal variabilities, impact of altitude variations, vehicle tilt or rotation, and inference speed of each method. Experimental results reveal that the YOLOv8 coupled with SAM and the YOLOv8-OBB exhibit comparable precision and excel in accurately localizing road users while preserving their dimensions. This can be instrumental in a practically feasible vision-based RTM solution. In terms of speed-to-error ratio, OBB-enabled object detectors present the most practical option, allowing for near-real-time solutions in key road safety workflows, such as conflict analysis.