Obstacles Avoidance Assistance for Visually Impaired
Description
This dataset was curated to support research on obstacle avoidance systems for visually impaired individuals. It consists of annotated images representing four common obstacle categories encountered by pedestrians like pole, fence, bump, and hole. The dataset was compiled from publicly available sources on Roboflow and refined with careful manual labeling in the YOLO format. Each image includes bounding box annotations corresponding to the obstacle class. The images were collected under various lighting conditions (daylight, low-light, and shadows) and different outdoor environments such as pavements, roads, and pedestrian walkways. The dataset contains a total of 1,627 annotations, distributed across classes as follows: - Pole: 361 annotations - Fence: 368 annotations - Bump: 479 annotations - Hole: 419 annotations
Files
Steps to reproduce
1. Data Collection - Images were sourced from publicly available datasets on Roboflow that contained outdoor pedestrian environments. - Selection focused on scenes with obstacles commonly encountered by visually impaired pedestrians, specifically poles, fences, bumps, and holes. 2. Data Curation - Images were filtered to remove duplicates and irrelevant backgrounds. - Care was taken to ensure variety in lighting conditions (daylight, low light, shadow) and environments (roads, pavements, pedestrian walkways). 3. Annotation Protocol - Each image was manually annotated in YOLO format using Roboflow’s labeling tool. - Bounding boxes were drawn tightly around each obstacle to reduce background noise and improve model accuracy. - Class labels were restricted to four categories: pole, fence, bump, hole. 4. Dataset Organization - The dataset was split into 80% training and 20% validation to support supervised learning. 5. Software and Tools - Roboflow was used for dataset preparation, annotation, and automatic formatting into YOLO format. - Ultralytics YOLOv8 framework (Python & Google Colab) was later used for training and validation of the model on this dataset.
Institutions
- Multimedia University