Application Of Lightweight Network In UAV Real-time Detection Of Ground Objects

Ground objects photographed by UAVs have typical small object characteristics.For example,the pixel area occupied by the object is small,and the edge information is blurred and difficult to distinguish from the background.These characteristics lead to relatively low detection precision and difficulty in improving.Therefore,improving the detection precision of small objects is one of the problems that must be solved for UAV vision to realize scene-based applications.This paper uses Vis Drone dataset and yolov5 s lightweight network model.In order to improve the small object detection precision of the model,focusing on the attention mechanism,low-level feature reuse,and bounding box loss function,the original yolov5 s model is improved as follows:(1)It is proposed to embed the attention module into the backbone network of yolov5 s,learn the correlation between each channel through a one-dimensional vector with the same number of channels as the feature map,and then weight the one-dimensional vector to the input feature map by channel multiplication.Recalibrate the original features in the channel dimension.(2)The low-level feature reuse is proposed.In the feature fusion stage of the neural network,the shallow P2 feature level with richer small object information is used,and a bottom-up feature fusion structure is added on the basis of FPN to shorten the information on the low-level feature map.The flow path in the entire feature fusion network retains more small object features and semantic information,thereby increasing the detection precision of small objects.(3)It is proposed to replace the loss function from GIoU with CIoU.CIoU introduces penalty items related to the distance between the center point of the prediction box and the groung truth box and the aspect ratio,which solves the degradation problem when the prediction box and the groung truth box completely overlap,by increasing the contribution of the loss of the small objects in the back propagation,the positioning of the prediction box is more accurate.Under the same experimental conditions,the results verify that the pricision of these three methods on the Vis Drone dataset is improved to a certain extent compared with the original yolov5 s model.When the three improved methods are integrated into the yolov5 s model,the pricision of small object detection has increased from 12.7% to37.66%,which is nearly three times higher.