Convolutional neural networks on small unmanned aerial systems

In the revolutionary field of deep learning many difficult computer vision challenges of today are being impressively overcome by the application of these cutting edge technologies. The challenge of detecting vehicular objects from aerial imagery is a long-standing interest in computer vision. Performing this accurately in real-time while utilizing the payload bay of a small unmanned aerial system (SUAS) is even more desirable and challenging. In this work, these challenges are successfully surmounted with the use of Faster R-CNN [1], a highly cultivated aerial image dataset, supercomputers, and a dedicated team of SUAS experts. By first training Faster R-CNN on a customized dataset of electro-optical (EO) annotated aerial imagery then empirically testing on supercomputers across hundreds of hyperparameters, the resulting optimized network was successfully integrated into established SUAS operations. This combination of cutting edge technologies lead to strong performance while requiring a fraction of the development time and meeting strict in-flight SWaP requirements.