DATA AUGMENTATION FOR CNN-BASED PEOPLE DETECTION IN AERIAL IMAGES

Abstract

Much more than ever, many important places have deployed surveillance cameras for early detection of abnormal events and suspects. However, the monitoring ability of fixed cameras is significantly limited due to the low flexibility, blind spot, and obstacle occlusion. With high mobility, drones have high potential for supporting security surveillance. On the other hand, people detection plays a key role in intelligent surveillance system, and increasing deep learning-based methods show great results. However, the training data for aerial images are still few, even though there are many public datasets available. Thus, in this paper we research on data augmentation, try transforming general images to be aerial image-like, and make an attempt to improving the performance of deep learning-based people detection with existing datasets. The experiments conducted on the real aerial images collected by a camera drone show encouraging results.