HazmatFireRescue Computer Vision Project

HazmatFireRescue: An Integrated Machine Learning Object Detection Dataset Tailored for RoboCup Rescue League

We are pleased to present DeepFireRescue, a refined object detection dataset purposefully tailored for developing machine learning models for the RoboCup Rescue League. This dataset amalgamates the strengths of two pre-existing datasets - DeepHAZMAT and FireNet - while removing specific classes to better cater to the requirements of RoboCup scenarios.

DeepHAZMAT, a dataset produced by Amir Sharifi, Ahmadreza Zibaei, and Mahdi Rezaei, was built to assist the detection of hazardous materials (HAZMAT) signs in low computational resource settings, specifically for robotic applications (Sharifi, Zibaei, & Rezaei, 2021).

On the other hand, FireNet, developed by J. Boehm, F. Panella, and V. Melatti at University College London, serves the purpose of fire detection, which is a crucial aspect in any rescue scenario (Boehm, Panella, & Melatti, 2019).

During the creation of the DeepFireRescue dataset, we carefully selected and combined the relevant classes from both datasets, optimizing for the context of the RoboCup Rescue League. Certain classes from the original datasets - specifically, White_Domes, Sounders, and Flashing_Light_Orbs - were deemed unnecessary for this particular application and hence excluded from DeepFireRescue.

We made diligent efforts to improve the quality of the FireNet dataset before integrating it with DeepHAZMAT. We resized some annotation bounding boxes, eliminated errors, and split bounding boxes that contained multiple object instances into individual bounding boxes with one object instance each. In addition, we expanded the DeepHAZMAT dataset by annotating the FireNet classes within its images where relevant examples were found.

Citations:

J. Boehm, F. Panella, and V. Melatti, “FireNet”. University College London, 31-Jul-2019, doi: 10.5522/04/9137798.v1.

Amir Sharifi, Ahmadreza Zibaei, and Mahdi Rezaei, "A deep learning based hazardous materials (HAZMAT) sign detection robot with restricted computational resources", Machine Learning with Applications, 2021, doi: https://doi.org/10.1016/j.mlwa.2021.100104.

Our aim with DeepFireRescue is to offer an efficient, streamlined dataset that is conducive to the specific needs of the RoboCup Rescue League. We trust this dataset will prove to be a beneficial tool for researchers and developers working towards advancing the field of robotic rescue missions.