弱光场景下锥套特征提取方法OA

Nighttime autonomous aerial refueling is a critical technology for enhancing the all-condition mission capability of unmanned aerial vehicles.To achieve real-time drogue positioning in low-illumination scenarios,an improved YOLO-based localization feature extraction method is proposed.Building upon the YOLO11s-oriented bounding box(OBB)rotation target detection network,this method enables direct elliptical feature regression independent of edge information.By introducing conditionally parameterized convolution(CondConv)to reconstruct the C3K2 feature extraction module,the convolutional kernel dynamically adjusts parameter weights based on input features,enhancing the model's feature generalization across different shadow situations,noise interference and target pose changes.A scale-sensitive elliptical regression(SSER)loss function is proposed.This function incorporates a scale-sensitive mechanism to strengthen the model's focus on dimensional information,effectively controlling depth estimation errors during drogue positioning.Experimental results demonstrate that drogue features extracted by our method achieve a mean positioning error of 126.3 mm and a standard deviation of 72.3 mm within 10 m,representing reductions of 21.5%and 41.7%,respectively,compared to the baseline model.Meanwhile,the average processing time per frame is 15.5 ms,meeting the requirements for nighttime autonomous aerial refueling drogue positioning and providing visual navigation support for autonomous nighttime aerial refueling.