This full dataset consists of 30,712 annotated images of a person wearing a green hazard vest in different outdoor scenarios of a university campus (IISc Bangalore). Our motivation for collecting this dataset is to improve the accuracy of detections of a person wearing a hazard vest and be able to track them autonomously using a drone. This is specifically designed to assist a Visually Impaired Person (VIP) navigate within urban spaces, as part of the Ocularone project described here.
The images are taken by a DJI Ryze Tello drone with an onboard camera. The camera has a Field of View (FoV) of 82.6 degrees and captures videos at a resolution of 720p at 30 frames per second (FPS).
A person holding the drone at different heights and distances from the person wearing the hazard vest, captures the videos. The frames are extracted from the videos at 10 FPS.
We have a total of 43 videos of duration between 1 minute to 2 minutes at different locations in the campus. The images have been manually annotated using Roboflow. Overall, we have 30,712 images in different scenarios as explained in the following table.
| Sl. No. | Scenarios | Number of annotated images |
|---|---|---|
| 1 | VIP walks on a footpath with usual surroundings | 2115 |
| 2 | VIP walks on a footpath with no pedestrians | 2294 |
| 3 | VIP walks on a footpath with pedestrians in the FoV | 1371 |
| 4 | VIP walks on a path with bicycles in the FoV | 901 |
| 5 | VIP walks on a path with pedestrians in the FoV | 1658 |
| 6 | VIP walks on a path with pedestrians and cycles in the FoV | 1057 |
| 7 | VIP walks on the side of road with pedestrians in the FoV | 1326 |
| 8 | VIP walks on the side of the road with usual surroundings | 1887 |
| 9 | VIP walks on the side of the road with no pedestrians in the FoV | 2022 |
| 10 | VIP walks on the side of the road with parked cars in the FoV | 2527 |
| 11 | VIP walks in broad daylight in different scenarios apart from the above specific ones | 9169 |
| 12 | Adversarial pictures of the VIP walking in dark backgrounds, blur pictures, bad orientations, evening time, etc. | 4384 |
A subset of these samples images have been released at this time (2024-03-15). The entire dataset will be shared once the relevant paper, currently under review, is accepted.
In these datasets, the person wearing the hazard vest is/was a team member of the project who has consented to be part of this data collection effort. They are not visually impaired; they just serve as a proxy for one. To respect privacy, we have blurred the faces of the person wearing the hazard vest and any other bystander whose features are recognizable.
We use 10% of images randomly picked from each category to train the off-the-shelf YOLOv8 nano Deep Neural Network (DNN) model. So, a total of 3866 images were set aside from the dataset and split in the ratio of 80:20 for training and validation data. Rest of the images were kept aside for testing.
- Suman Raj
- Bhavani A Madhabhavi
- Prince Modi
- Arnav Rajesh
- Pratham M
- Yogesh Simmhan
You may cite this work as follows:
Ocularone Hazard Vest Dataset, v1.0, Suman Raj, Bhavani A Madhabhavi, Prince Modi, Arnav Rajesh, Pratham M, Yogesh Simmhan, DREAM:Lab, Indian Institute of Science, Bangalore, 2024, https://github.com/dream-lab/ocularone-dataset
Ocularone Hazard Vest Dataset by DREAM:Lab, Indian Institute of Science, Bangalore is licensed under Attribution-NonCommercial 4.0 International
© Indian Institute of Science, Bangalore, 2024