Abnormal road objects detection algorithm in UAV aerial videos based on improved YOLOv5s
During the use of UAV for pedestrian and non-motorized vehicle detection in motorway,the problem of low accuracy and poor performance in object detection was found.To solve these prob-lems,a pedestrian and non-motorized vehicle detection algorithm YOLOv5s-P2S was proposed for UAV perspective.Firstly,the neck part of the YOLOv5s model was extended based on the original PAFPN feature fusion scheme,and a detection layer specifically for small object was added.Then,a small object detection head was added to the prediction part to predict the output feature map of the small object detection layer.Finally,the localization loss function of YOLOv5s was modified to SIOU to improve the detection accuracy and regression efficiency of the anchor box.The experimental results showed that compared with the YOLOv5s model,the average accuracy mean mAP50 of YOLOv5s-P2S increased by 0.05,and the parameters only increased by 0.2M.YOLOv5s-P2S can meet the accuracy and real-time requirements of pedestrian and non-motorized vehicle object detection for UAV perspective.
国家科技期刊平台
NSTL
万方数据
