Aerial Vehicle And Pedestrian Detection Algorithm Based On Multi-Scale And Dense Feature Fusion

With the development of drone technology,the aerial image of drone is also applied to the field of object detection as a new image.Vehicle and pedestrian detection in drone aerial images plays a positive role in the research and development of intelligent transportation.Deep learning method is one of the popular methods for studying object detection nowadays,and it has great advantages compared with traditional object detection methods.Therefore,this paper chooses a deep learning-based object detection algorithm to study vehicle and pedestrian detection in aerial images.First of all,in order to solve the problems of small-scale aerial objects,severe occlusion,complex scenes,and real-time detection,an improved YOLOv3algorithm--H-YOLOv3 algorithm is proposed.H-YOLOv3 algorithm for the problem that the multi-scale detection anchor box designed by YOLOv3 algorithm is not suitable for aerial small object detection,redesigned the anchor box in line with aerial small object detection.The H-YOLOv3 algorithm aims at the problem that YOLOv3 algorithm is not suitable for aerial small object detection of low-resolution multi-scale feature map.Four different resolution feature maps extracted by feature extraction network of YOLOv3 algorithm are analyzed for receptive field.According to the matching relationship between receptive field and object to be detected,high-resolution feature map is selected for multi-scale detection.In view of the problem that the feature pyramid structure of the YOLOv3 algorithm does not perform the splicing operation of different resolution features for the feature map at the top of the pyramid,which is not conducive to aerial small object detection,the H-YOLOv3 algorithm designs high resolution two stages of re-sampling feature pyramid structure,i.e.up sampling and down sampling,to improve the feature pyramid structure of the YOLOv3 algorithm.Then,an improved H-YOLOv3 algorithm--DH-YOLOv3 algorithm,is proposed to solve the problem of missing detection of H-YOLOv3 algorithm on smaller targets such as aerial pedestrians.DH-YOLOv3 algorithm improves the network structure of feature extraction of H-YOLOv3 algorithm.On the basis of not increasing the amount of network training parameters,it uses the method of dense feature fusion to integrate the features rich in details in the shallow layer of the feature extraction network with the features rich in semantics in the deep layer,fully retaining the feature information of smaller targets and forming the feature map which is good for small target detection in aerial photographing,and alleviating the problem of missing detection of human small target.Finally,using the visDrone aerial dataset,under the same experimental environment,multiple sets of comparative experiments were set up to verify the improved method;the DH-YOLOv3 algorithm was compared with the currently popular object detection algorithms for detection accuracy comparison and detection speed comparison.