IEEE Access (Jan 2024)
ATR HarmoniSAR: A System for Enhancing Victim Detection in Robot-Assisted Disaster Scenarios
Abstract
As our world increasingly faces various disastrous events, there is an urgent need for improved methods of disaster response, particularly for the detection of victims trapped in the debris. Deep learning presents a promising avenue for this task. However, the applicability of advanced detection models trained on popular datasets like COCO is limited, given their focus on objects in unobstructed, everyday conditions. In stark contrast, victims in disaster scenarios are often buried or partially hidden by rubble, presenting a unique challenge for detection models. Additionally, the collection of real-world disaster victim images for training is a daunting and ethically complex task. This study seeks to address these challenges by proposing an all-in-one solution for generating realistic, synthetic images of disaster victims using a framework based on the Poisson equation. This harmonious composite image generation provides a versatile and accessible means of training deep learning models for victim detection, circumventing the difficulties associated with gathering natural disaster images. We leverage the YOLO architecture for training and testing our model, applying it to the synthesized harmonious images. Our approach seeks to balance the need for effective, reliable victim detection in robot-assisted search and rescue missions with the practical and ethical constraints of model training. By generating synthetic yet realistic representations of disaster scenarios, we aim to create a solution that generalizes well and thus can potentially enhance the efficacy of disaster response efforts. We have achieved the best Average Precision (AP) with 50% Intersection over Union (IoU) called ( $AP_{50}$ ) using YOLOv5x. The ( $AP_{50}$ ) for YOLOv5x after just training was 29.3% but it reached 92.4% after harmonization.
Keywords
