There are many advantages in using UAVs for search and rescue operations. However, detecting people from a UAV remains a challenge: the embedded detector has to be fast enough and viewpoint robust to detect people in a flexible manner from aerial views. In this paper we propose a processing pipeline to 1) reduce the search space using infrared images and to 2) detect people whatever the roll and pitch angles of the UAV's acquisition system. We tested our approach on a multimodal aerial view dataset and showed that it outperforms the Integral Channel Features (ICF) detector in this context. Moreover, this approach allows real-time compatible detection.