Automated Inspection of Obstructed Fire Extinguishers Using Amodal Instance Segmentation

Fire safety inspections are essential to ensure the safety of occupants during a fire outbreak. These inspections involve checking various items, such as fire safety equipment (FSE), to ensure they are unobstructed and fully functional. Recent research has highlighted the potential of machine learning and computer vision in automating and enhancing fire safety inspection processes using image analysis. However, identifying obstacles or potential obstructions in front of fire extinguishers effectively remains a challenge. The research focuses on reviewing documented images to identify instances where extinguishers are obstructed, either partially or fully. This study proposes a novel approach to address this challenge by combining modal and amodal instance segmentation models to evaluate the level of obstruction. One conducts classical (modal) instance segmentation of visible fire extinguisher parts, while the other performs amodal segmentation. Additionally, the annotation and dataset creation for amodal segmentation tasks is addressed. The study generates obstacles on images from both open-source and self-created datasets containing (unblocked) fire extinguishers. This approach requires only one modal annotation iteration to generate modal and amodal annotation data. Results demonstrate the effectiveness of the proposed approach in detecting covered or partially blocked extinguishers. Future research aims to refine amodal mask results and extend the approach to other FSE components, further enhancing fire safety inspection processes.