ROD-Dataset: Real-Time Obstacle Detection for Smartphone-Based Assistive Vision

24,326-image, 25-class YOLO dataset for obstacle detection

This dataset is the data product of our Real-Time Obstacle Detection (ROD) project at Amirkabir University of Technology, Tehran. The project addresses two related public-safety problems on the city sidewalk: the limited situational awareness of people living with visual impairments, and the elevated collision and fall risk for pedestrians who walk while looking at their phones. The deployed system runs an optimized YOLOv8n detector directly on a mid-range Android phone, pairs it with ARCore for monocular distance estimation, and delivers feedback as Text-to-Speech for visually impaired users and as vibration cues for distracted ones. The whole pipeline is built to run on consumer hardware, so no LiDAR, depth camera, or external sensor is required.

The release contains 24,326 annotated images and 40,195 bounding boxes across 25 obstacle categories, split 19,186 / 3,511 / 1,629 into train, validation, and test. Annotations follow the standard YOLO Darknet TXT format (one line per box, normalized coordinates), and the class-index mapping is fixed in data.yaml. The 25 classes cover vehicles (Car, Bus, Truck, Motorcycle, Bike), street users (Person, Dog), built-environment elements (Building, Tree, Stairs, Manhole, Guard rail, Pedestrian crosswalk, Road), and the kinds of street furniture that general-purpose detectors typically miss in practice (Dustb