Battle of the YOLOs - Augmenting and Benchmarking Fisheye Camera YOLO Models for Traffic Intersection Object Detection
DOI:
https://doi.org/10.32473/ufjur.27.139103Keywords:
computer vision, high-performance computing, neural networks, pedestrian safetyAbstract
Our data pipeline, which leverages a computer vision model, can ingest fisheye camera footage and compute statistics such as car volume and the near miss rate per intersection. However, accurate metrics are dependent on the model’s class detection accuracy. As new versions of Darknet/YOLO (You Only Look Once) are released and changes to the intersection are completed, the model must be retrained and benchmarked to improve accuracy. We report on a new, integrated process that automates model training from annotation to deployment, producing a YOLOv4 model with 99.10% mAP (mean average precision) across all classes within a fisheye traffic camera environment. The training is containerized and portable across computer environments, including consumer-grade personal computers and high-performance computing clusters.
Accessibility Summary:
In accordance with Title II regulations this content meets all points of exemption as Archived web content and/or Preexisting conventional electronic documents.
Downloads
Published
Issue
Section
License
Copyright (c) 2025 Jacques Fleischer, Julian Dominguez, Artur Kumik, Sargam Thakur, Gregor von Laszewski, Sanjay Ranka
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Some journals stipulate that submitted articles cannot be under consideration for publication or published in another journal. The student-author and mentor have the option of determining which journal the paper will be submitted to first. UF JUR accepts papers that have been published in other journals or might be published in the future. It is the responsibility of the student-author and mentor to determine whether another journal will accept a paper that has been published in UF JUR.
