Optimizing High-Altitude UAV Object Detection with Deep Learning

Abstract

The impressive advancements in computer vision have endowed unmanned aerial vehicles (UAVs) with intelligence, enabling them to undertake diverse tasks such as object detection, tracking, and counting. Nevertheless, challenges arise due to factors like high altitudes, varying camera angles, and environmental conditions. UAVs are often equipped with thermal cameras, offering valuable footage for a broad spectrum of applications. In this study, we focused on a high-Altitude infrared thermal dataset, investigating different models within the YOLOX family for the object classification task. The dataset comprises 2898 images taken at various times, altitudes, weather conditions, and camera perspectives. Notably, the YOLOX-L model achieved the highest mean average precision (mAP) at 86.5. In comparison, previous work utilizing SSD-512 achieved an mAP of 85.6, demonstrating our improved precision. It is noteworthy that even the compact models within the YOLOX family yielded commendable precision. YOLOX-Tiny achieved 84.5, and YOLOX-nano achieved 82.8. In contrast, prior research indicated that the YOLOv4-Tiny model only achieved an mAP of 50.38. Considering the computational complexity, the YOLOX tiny models prove highly advantageous for deployment in onboard UAV platforms.