In the one-stage object detection based on deep learning,the bounding box regression loss based on the Intersection over Union (IoU) is not sensitive enough to the change of the position relationship of the bounding box.The existing loss cannot accurately distinguish different inclusion relationships between the predicted frame and the true value frame.In response to the above problem,a loss based on Regression Position Relationship Sensitivity IoU (RPIoU) is proposed.This loss design can strengthen the sensitivity to the relative positional relationship between the predicted frame and the true value frame.Firstly,a penalty term is added after the IoU to make the corners of the two frames infinitely close.It solves the problem of IoU degradation when the center points coincide.Secondly,the exponential function taking the ratio of the area of non-overlapping region to that of the true value frame as the parameter is introduced as the penalty term,which can deal with the problem that the loss cannot distinguish different inclusion relationships between the predicted frame and the true value frame,and can guide the position of frame regression more accurately.Considering that each part of the total loss of the one-stage object detection algorithm has different degrees of contribution to the training results,this paper takes the mean Average Precision (mAP) as the fitness function,and uses the genetic algorithm to optimize the total loss of training to obtain the optimal weights of classification,regression and confidence loss respectively.The designed loss is applied to the one-stage object detection algorithm YOLOv5,which is verified on the public visible light dataset VisDrone and the self-made infrared aircraft dataset respectively.On the public visible light dataset,the mAP reaches 0.447,which is 0.037 higher than that of the original YOLOv5.On the infrared aircraft dataset,the mAP reaches 0.966,which is 0.014 higher than that of the original YOLOv5.