The combination of unmanned aerial vehicles (UAVs) and deep learning has potential applicability in various complex search and rescue scenes. However, due to the presence of environmental occlusions such as trees, the performance of UAVs mounted with different optical payloads in detecting missing persons is poor. To the best of our knowledge, currently available non-occluded human target datasets are insufficient to address the challenges of automatic recognition for partially occluded human targets. To address this problem, we collected a UAV-based infrared thermal imaging dataset for outdoor, partially occluded person detection (POP). POP is composed of 8768 labeled thermal images collected from different environmental scenes. After training with popular object detection networks, our dataset performed stable average precision for partially occluded person detection and short response time. In addition, high precision of object detection by POP trained networks was not attenuated until the occlusion rate exceeded 70%. We expected POP would extend present methodologies for the search of human objects under complex occluded circumstances.