Deep learning-based drone detection in infrared imagery with limited training data

The increased availability and capabilities of drones in the consumer market has lead to increased risk in air traffic control and other public safety concerns. Automated drone detection systems can help to generate alerts and increase reaction time by security forces. Recently proposed approaches and systems are usually based on a combination of sensors and machine learning to carry out the detection of drones. While electro-optical imagery is the most prevalent modality, infrared sensors can complement it by providing better visibility in certain situations with cluttered background or low light conditions. A key limitation when using infrared data is the limited availability of data for training machine learning methods. In this work, we specifically focus on the task of drone detection in infrared imagery. Our main focus lies on investigating how the small amount of available infrared data can be compensated for. We approach this problem through three different types of experiments. First, we compare a detector resulting from training on limited infrared data with a detector trained on more diverse optical data. We then propose and evaluate several methods for pre-processing optical data in such a way that it better resembles the characteristics of infrared data. Finally, we train detectors on a combination of infrared and pre-processed optical data and evaluate the trade-off between amount of available infrared data and achieved accuracy of the resulting detector. We evaluate all detectors on our own set of diverse infrared recordings. Our results show that suitable pre-processing of optical data can significantly improve the resulting accuracy and performs much better than training solely on limited infrared data.