Large Area Building Detection from Airborne Lidar Data using OSM Trained Superpixel Classification

Lidar-based building detection has been an interesting topic for a long time considering its wide applications such as change detection, model reconstruction and so on. In this paper, we propose a Lidar building detection algorithm based on superpixel segmentation. Compared with the existing methods, two innovations of this paper are first a deep learning model is trained for classification of the superpixel using OpenStreetMap as semantic tags. Second, we further improve our segmentation results by graph analysis of the generated superpixels. The experimental results indicate that a simple CNN based deep learning model trained in the proposed framework can reach over 80% classification accuracy of building objects. And after the graph analysis, the building detection accuracy can further increase to 98%. According to the tests in the London area, the proposed method can be used for large-scale model detection with low-resolution Lidar data (50CM).