Training of YOLO Neural Network for the Detection of Fire Emergency Assets

Building assets surveys are cost and time demanding and the majority of current methods still rely on manual procedures. New technologies could be used to support this task. The exploitation of Artificial Intelligence (AI) for the automatic interpretation of data is spreading throughout various application fields. However, a challenge with AI is the very large number of training images required for robustly detect and classify each object class. This paper details the procedure and parameters used for the training of a custom YOLO neural network for the recognition of fire emergency assets. The minimum number of pictures for obtaining good recognition performances and the image augmentation process have been investigated. In the end, it was found that fire extinguishers and emergency signs are reasonably detected and their position inside the pictures accurately evaluated. The use case proposed in this paper for the use of custom YOLO is the retrieval of as-is information for existing buildings. The trained neural networks are part of a system that makes use of Augmented Reality devices for capturing pictures and for visualizing the results directly on site.