Aprus: An Airborne Altitude-Adaptive Purpose-Related UAV System for Object Detection

Unmanned aerial vehicle (UAV) system plays an important role in new edge scenarios such as disaster relief, situational awareness, and so on. In order to improve efficiency, computing tasks such as object detection are carried out on the drone-mounted computing unit. However, there is a trade-off between flight altitude and mission accuracy. Higher altitudes result in smaller images, which influence the accuracy. In comparison, shorter altitudes reduce the difficulty of detection while decreasing the recognition range, which affects user experience. Therefore, this paper proposes an airborne altitude adaptive UAV System, Aprus, which selects the most suitable object detection models based on a self-adaptation strategy at different altitudes. The selection strategy is based on a purpose-related evaluation indicator, PEI. It comprehensively considers the model's accuracy, recall, and inference speed at the current altitude according to the user's purpose. Moreover, Aprus sends the original image to a divider instead of scaling the image before pushing it to the model, thus ensuring that the original picture information will not be lost. To evaluate the system, we build a high-resolution UAV dataset with altitude, UDWA, which contains 46037 images. From the experiments, Aprus obtained 58.52% mAP, 94.17%mAP ⁵⁰ , 66.17%mAR, 1.61FPS results on DJI Manifold 2G when the purpose is set to Balance mode, and the system can be adjusted by multiple preset purposes or according to the user customs.