How to improve the HOG detector in the UAV context

The well known HOG (Histogram of Oriented Gradients) of Dalal and Triggs is commonly used for pedestrian detection with 2d moving embedded cameras (driving assistance) or static cameras (video surveillance). In this paper we show how to use and improve the HOG detector in the UAV context. In order to increase the elevation angular robustness we propose to use a more appropriate training dataset and sliding windows. We show results on synthetic images.