| The use of unmanned aerial vehicles(UAVs)in remote sensing and surveillance applications has increased dramatically in recent years,and research on visual navigation,object detection and path planning is becoming a hot topic.UAVs offer multiple advantages and have tremendous economic benefits in many areas.Rapidly developing deep learning techniques have made remarkable progress in computer vision tasks,effectively reducing the over-reliance of classical feature extraction algorithms on artificially designed features,and therefore,it is of great research significance and economic value to study object detection methods for UAV aerial images based on deep learning techniques.However,it is challenging to design a high-accuracy UAV aerial image object detection algorithm because of the characteristics of UAV aerial images such as severe object occlusion,more small objects and complex image backgrounds.In this thesis,we use deep learning to fuse occlusion information,increase the proportion of low-level features and align features to cope with these problems in UAV aerial photography images,and the main research contents and results are as follows:(1)The improved DDETR object detection algorithm incorporating occlusion information is proposed to cope with the difficult object detection and small object miss detection problems in the occlusion case of object detection.An occlusion degree estimation module is proposed to assist the model in solving the occlusion problem,and an occlusion loss function matching the occlusion prediction task is designed to enable the model to better detect heavily occluded objects by evaluating the object’s occlusion degree.A feature mapping module with more low-level features is designed to improve the detection of small and medium-sized objects.Several sets of experimental results show that the proposed method has better detection results.(2)The improved DDETR object detection algorithm with feature enhancement and alignment is proposed to solve the problem of difficult detection of small objects in complex backgrounds.In the framework of DDETR algorithm,the feature alignment pyramid is used to learn the semantic information offset during feature fusion,so as to align multi-level features and improve the detection ability of the model for small objects.To address the problem of difficult object detection in complex background images,the Swin Transformer is used instead of the residual network in the DDETR model to model complex scenes and extract multi-level features with richer semantic information,so as to better utilize the semantic information in images to detect objects.Experiments show that the improved DDETR algorithm with feature semantic information enhancement and alignment has higher object detection accuracy and reaches an advanced level compared with other mainstream methods. |
