Core Concepts
This study presents a comprehensive benchmark of 22 state-of-the-art YOLO object detectors for real-time traffic object detection, with a focus on enhancing the safety of electric scooters.
Abstract
The key highlights and insights from the content are:
The study addresses the gap in research on applying deep learning-based object detection to enhance the safety of electric scooters (e-scooters), which have seen a concerning rise in related injuries and fatalities.
A comprehensive benchmark involving 22 YOLO object detectors, including five versions (YOLOv3, YOLOv5, YOLOv6, YOLOv7, and YOLOv8), was established using a self-collected dataset featuring e-scooters and other traffic objects.
The detection accuracy, measured in terms of mAP@0.5, ranged from 27.4% (YOLOv7-E6E) to 86.8% (YOLOv5s). All YOLO models, particularly YOLOv3-tiny, displayed promising potential for real-time object detection in the context of e-scooters.
The study analyzed the trade-off between the effectiveness (detection accuracy) and efficiency (model complexity and inference time) of the YOLO detectors. YOLOv5 variants, especially YOLOv5n and YOLOv5s, as well as YOLOv3-tiny, were highlighted for their superior computational efficiency and swift inference times.
The traffic scene dataset (ScooterDet) and software program codes for model benchmarking were made publicly available, which will not only improve e-scooter safety with advanced object detection but also lay the groundwork for tailored solutions, promising a safer and more sustainable urban micromobility landscape.
Stats
The ScooterDet dataset contains 2,013 images with a total of 11,011 bounding box annotations across 11 traffic object classes.
YOLOv5s achieved the highest mAP@0.5 of 86.8% among the evaluated models.
YOLOv5n and YOLOv5s had the fastest inference times, under 5 milliseconds.
Quotes
"All YOLO models, particularly YOLOv3-tiny, have displayed promising potential for real-time object detection in the context of e-scooters."
"YOLOv5 variants—specifically YOLOv5n and YOLOv5s—and YOLOv3-tiny are highlighted for their superior computational efficiency and swift inference times (under 5 milliseconds)."