WalkSentry

← All research

Under review2026

ROD-Dataset: A Pedestrian-Perspective Benchmark of 25 Urban Obstacle Classes for Real-Time Detection

Abtin Zandi, Ariyan Azami, Bardia Sabbagh Kermani, Parsa Abbasian, Roza Ganjipour, Sarvin Nami, Hamed Farbeh

  • Paper
  • arXiv
  • DOI
  • Code
  • Dataset

Abstract

ROD-Dataset is a pedestrian-perspective benchmark for real-time obstacle detection. It contains more than 24,000 eye-level images and over 40,000 human-verified obstacle annotations across 25 classes, from vehicles and people to stairs, manholes, traffic cones and benches. It unifies 29 public collections under a single taxonomy and adds new captures from Tehran and Toronto. We benchmark six nano-scale YOLO detectors, all of which exceed 0.86 recall: YOLO12n achieves the best recall (0.889), YOLO26n the best precision (0.925), and YOLOv9t the best mAP@0.50 (0.882).

Highlights

  • 24,000+ eye-level images and 40,000+ human-verified obstacle annotations.
  • 25 obstacle classes, from vehicles and people to stairs, manholes, traffic cones and benches.
  • 29 public collections unified onto one taxonomy, plus new captures from Tehran and Toronto.

Baseline results

Six nano-scale YOLO detectors were benchmarked; all exceed 0.86 recall.

Detector Best at Score
YOLO12n Recall 0.889
YOLO26n Precision 0.925
YOLOv9t mAP@0.50 0.882

Cite this work

@misc{zandi2026roddataset,
  title  = {ROD-Dataset: A Pedestrian-Perspective Benchmark of 25 Urban Obstacle Classes for Real-Time Detection},
  author = {Zandi, Abtin and Azami, Ariyan and Sabbagh Kermani, Bardia and Abbasian, Parsa and Ganjipour, Roza and Nami, Sarvin and Farbeh, Hamed},
  year   = {2026},
  note   = {Under review}
}