Rooftop Detection using Aerial Drone Imagery

We present a rooftop detection algorithm using aerial RGBD and near infrared data which uses lower computational resources than algorithms requiring GPUs. The depth data is extracted from multi-view images captured by drones using photogrammetry. Our approach is cost effective as compared to LIDAR surveying and has lower edge blurring. It is also novel in terms of segmenting clutter due to objects such as overhead water tanks on roofs. This helps is determining the actual free roof area that would be available for applications like solar panel deployment. The algorithm was evaluated on the aerial imagery of rooftops on a hill slope of size 12876x10533 pixels and an F1 score of 88.27% was obtained. The algorithm ran in under 2 minutes on a Google Cloud instance with Intel Xeon E5 processor.