Aerial Orthophoto Target Recognition and Positioning Method based on Improved YOLOX

The ground target recognition and localization method relying on UAV vision sensors can avoid carrying complex sensing equipment and reduce equipment costs. The improved deep learning YOLOX (You Only Look Once version X) target detection algorithm is proposed to be applied to the aerial photography target recognition module, and the conversion model from pixel coordinates to geodesic coordinates based on orthorectified images is constructed in the target localization module, and the input images are rotated orthorectified with the help of UAV heading angle information. Based on the fixed-wing UAV and Pytorch deep learning framework, the fast target recognition and localization of ground vehicles in aerial photography environment is simulated and realized by making aerial photography vehicle dataset, and its target recognition mAP value reaches 90%, and localization accuracy reaches less than 20 meters with real-time. The method can provide some reference for certain application scenarios that do not require high localization accuracy.