Due to now outdated construction technology, houses which have not been retrofitted since construction typically fail tomeet modern energy performance levels. However, identifying at a city scale which houses could benefit the most fromretrofit solutions is currently a labour intensive process. In this paper, a system that uses a vehicle mounted camera tocapture pictures of residential buildings and then performs semantic segmentation to differentiate components of capturedbuildings is presented. An ensemble of U-Net semantic segmentation models are trained to identify walls, roofs,chimneys, windows and doors from building fa?ade images and differentiate between window and door instances whichare partially visible or obscured. Results show that the ensemble of U-Net models achieved high accuracy in identifyingwalls, roofs and chimneys, moderate accuracy in identifying windows and low accuracy in identifying doors andinstances of windows and doors which were partially visible or obscured. When U-Net models were retrained to identifydoors or windows, irrespective of partially visible and obscured instances, a significant rise in door and windowidentification accuracy was observed. It is believed that a larger training dataset would produce significantly improvedresults across all classes. The results presented here prove the operational feasibility in the first part of a process tocombine this model with high-resolution thermography and GPS for automating building retrofitting evaluations.
展开▼
