ROD-Dataset: A Pedestrian-Perspective Benchmark of Urban Obstacle Classes for YOLO-Based Detection
Organizations: Department of Computer Engineering, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran. · York University, Toronto, Canada. · Department of Mathematics and Informatics, Al-Khwarizmi University, Urgench, 220100, Uzbekistan.
Abstract
Two groups of pedestrians face avoidable risk on the city sidewalk: people living with visual impairment, who reach an obstacle before anything warns them, and people walking with divided attention, who register a hazard late or not at all. A vision system can warn them only about obstacles it has learned to recognize, so its training data limits the protection it offers. No suitable training set exists: public benchmarks are recorded from vehicles, aerial platforms, or fixed rigs, general-purpose detectors omit the street furniture that causes trips, and pedestrian-focused collections are confined to single regions with incompatible class schemes. This paper introduces the Real-time Obstacle Detection (ROD) dataset: more than 24,000 images and 40,000 annotated obstacles across 25 classes, from vehicles and people to the stairs, manholes, traffic cones, and benches that general-purpose detectors overlook. Because street furniture differs between cities, ROD unifies twenty-nine public collections onto one taxonomy and adds eye-level imagery from two contrasting cities on two continents, Tehran and Toronto. Every label is human-verified, and each capture campaign is admitted only after a compact detector shows it can learn the class. Six lightweight YOLO detectors (nano and tiny variants) benchmarked on ROD all exceed 0.86 recall; YOLO12n attains the highest recall (0.889), YOLO26n the highest precision (0.925), and YOLOv9t the best [email protected] (0.882) with the fewest parameters. Because the recall and mAP rankings disagree, selecting a model by mAP does not minimize missed obstacles, so per-class recall is reported for all 25 classes.
Figures & tables
| Dataset | Viewpoint | Images | Cls. | Scope |
|---|---|---|---|---|
| Waymo Open Sun et al. (2020) | Vehicle roof | 1,150 seq. | 4 | Driving |
| DOTA Ding et al. (2022) | Aerial | 11,268 | 18 | Remote sensing |
| FieldSAFE Kragh et al. (2017) | Tractor | 2 h | — | Agriculture |
| MaSTr1325 Bovcon et al. (2019) | USV deck | 1,325 | — | Marine |
| UAV-RSOD Rampriya et al. (2024) | UAV | 2,002 \tnote a | 6 | Railway |
| SVRDD Ren et al. (2024) | Street view | 8,000 | 7 | Pavement damage |
| Group | Images | Boxes | Train / valid / test images |
|---|---|---|---|
| Field captures, identified | 1,582 (6.5%) | 1,707 (4.2%) | 1,398 / 92 / 92 |
| All other images | 22,744 (93.5%) | 38,488 (95.8%) | 17,788 / 3,419 / 1,537 |
| Release | 24,326 | 40,195 | 19,186 / 3,511 / 1,629 |
| Class | Field boxes | Class total | Share |
|---|---|---|---|
| Fire hydrant | 250 | 255 | 98.0% |
| Bench | 514 | 774 | 66.4% |
| Electrical Pole | 159 | 618 | 25.7% |
| Dustbin | 376 | 1,549 | 24.3% |
| Traffic Cone | 209 | 941 | 22.2% |
| Stairs | 81 | 1,687 | 4.8% |
| Capture campaign | Precision | Recall | mAP 50 | mAP 50:95 |
|---|---|---|---|---|
| Fire hydrant | 0.997 | 1.000 | 0.995 | 0.917 |
| Traffic Cone | 0.996 | 1.000 | 0.995 | 0.905 |
| Electrical Pole | 0.985 | 1.000 | 0.995 | 0.724 |
| Manhole | 0.971 | 1.000 | 0.995 | 0.648 |
| Dustbin | 0.991 | 0.933 | 0.988 | 0.745 |
| Bench | 0.769 | 0.895 | 0.860 | 0.521 |
| Split | Images | Instances | Instances/Image |
|---|---|---|---|
| Train | 19,186 | 32,251 | 1.68 |
| Validation | 3,511 | 5,572 | 1.59 |
| Test | 1,629 | 2,372 | 1.46 |
| Total | 24,326 | 40,195 | 1.65 |
| ID | Class | Train | Val | Test | Total |
|---|---|---|---|---|---|
| 0 | Bike | 323 | 95 | 49 | 467 |
| 1 | Building | 886 | 101 | 54 | 1,041 |
| 2 | Car | 4,328 | 1,003 | 247 | 5,578 |
| 3 | Person | 7,609 | 1,033 | 320 | 8,962 |
| 4 | Stairs | 1,125 | 339 | 223 | 1,687 |
| 5 | Traffic sign | 689 | 114 | 104 | 907 |
| Model | Params (M) | P | R | mAP 50 | mAP 50:95 | ms | FPS |
|---|---|---|---|---|---|---|---|
| YOLOv8n | 3.01 | 0.908 | 0.870 | 0.876 | 0.718 | 12.5 | 79.7 |
| YOLOv9t | 1.98 | 0.904 | 0.880 | 0.882 | 0.750 | 14.2 | 70.4 |
| YOLOv10n | 2.27 | 0.914 | 0.867 | 0.871 | 0.726 | 13.5 | 74.2 |
| YOLO11n | 2.59 | 0.915 | 0.872 | 0.880 | 0.735 | 14.3 | 70.1 |
| YOLO12n | 2.56 | 0.899 | 0.889 | 0.878 | 0.739 | 16.7 | 59.7 |
| YOLO26n | 2.38 | 0.925 | 0.876 | 0.874 | 0.742 | 12.9 | 77.6 |
| Class | YOLOv8n | YOLOv9t | YOLOv10n | YOLO11n | YOLO12n | YOLO26n |
|---|---|---|---|---|---|---|
| Bike | 0.898 | 0.957 | 0.918 | 0.939 | 0.959 | 0.939 |
| Build. | 0.955 | 0.979 | 0.981 | 0.963 | 0.981 | 0.961 |
| Car | 0.820 | 0.814 | 0.794 | 0.810 | 0.842 | 0.834 |
| Person | 0.819 | 0.819 | 0.757 | 0.817 | 0.859 | 0.787 |
| Stairs | 0.888 | 0.901 | 0.798 | 0.871 | 0.907 | 0.892 |
| Sign | 0.933 | 0.865 | 0.840 | 0.885 | 0.879 | 0.885 |
| System / dataset | Cls. | Precision | Recall | mAP 50 |
|---|---|---|---|---|
| Obstacle Finder huday (2024) , outdoor | 15 | 0.736 | 0.584 | 0.632 |
| obstacle ThiGiacMayTinh (2024) , indoor | 9 | 0.797 | 0.809 | 0.846 |
| ROD, YOLO12n (proposed) | 25 | 0.899 | 0.889 | 0.878 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| # | Source | Roboflow project | License |
|---|---|---|---|
| 1 | Test (2024) | people-s773n | CC BY 4.0 |
| 2 | Hasan (2024) | stair-detection-project | CC BY 4.0 |
| 3 | Zlata (2024) | tree-classification-qbhpo | CC BY 4.0 |
| 4 | Project (2024) | road-users-detection-a2kmr | CC BY 4.0 |
| 5 | Titulacin (2025) | person-detection-9a6mk | CC BY 4.0 |
| 6 | tensorflow (2024) | minor_dataset-al08q | CC BY 4.0 |
Object Detection
Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.
See all 673 papers