| Popis: |
Accurate fire load (combustible objects) information is crucial for safety design and resilience assessment of buildings. Traditional fire load acquisition methods, such as fire load survey, which are time-consuming, tedious, and error-prone, failed to adapt to dynamic changed indoor scenes. As a starting point of automatic fire load estimation, fast recognition and detection of indoor fire load are important. Thus, this research proposes a computer vision-based method to automatically detect indoor fire loads using deep learning-based instance segmentation. First, indoor elements are classified into different categories according to their material composition. Next, an image dataset of indoor scenes with instance annotations is developed. Finally, a deep learning model, based on Mask R-CNN, is developed and trained using transfer learning to detect fire loads in images. Experimental results show that our model achieves promising accuracy, as measured by an average precision (AP) of 40.5% and AP50 of 59.2%, for instance segmentation on the dataset. A comparison with manual detection demonstrates the method’s high efficiency as it can detect fire load 1200 times faster than humans. This research contributes to the body of knowledge 1) a novel method of high accuracy and efficiency for automated fire load recognition in indoor environments based on instance segmentation; 2) training techniques for a deep learning model in a relatively small dataset of indoor images which includes complex scenes and a variety of instances; and 3) an image dataset with annotations of indoor fire loads. Although instance segmentation has been applied for several years, this is a pioneering research on using it for automated indoor fire load recognition, which paves the foundation for automatic fire load estimation and resilience assessment for the built environment. |
