Object Detection In Unmanned Aerial Vehicle Remote Sensing Images

With the development of satellite communication technology,spatial positioning tech-nology,remote sensing technology and geographic information systems,the demand and us-age of high-resolution remote sensing images from Unmanned Aerial Vehicles(UAV)are gradually increasing.As one of the typical applications of remote sensing images,object de-tection in UAV remote sensing images has always been a hot issue for researchers.As an important part of intelligent remote sensing information processing,it plays an important role in urban planning,environmental monitoring,animal protection,military target strikes,ab-normal situation analysis and other applications.However,due to the flexibility of UAV,the perspectives,background distributions and target scales are quite different on different im-ages,which increases the difficulty of processing UAV remote sensing images.Compared with natural scene images and other remote sensing images,there are more difficulties and challenges in object detection in UAV remote sensing images:(1)Due to the diverse perspec-tives of remote sensing images,objects in remote sensing images often have different postures and shapes.(2)The remote sensing sensor is relatively far from the object,which results to the larger map of remote sensing images and the complicated background situation of object location.(3)In addition,there are multi-scale objects in UAV remote sensing images,which brings great difficulty to object detection tasks.To solve the above problems,this thesis carries out a series of object detection methods by using the characteristics of UAV remote sensing images.The main work of this thesis is as follows:(1)Aiming at solving the difficulty of object detection due to the variable viewing an-gles in UAV remote sensing images,this thesis proposes a cross view adversarial network for object detection in UAV remote sensing images.This method uses a dual-branch adversarial network to alternately train single-view satellite images and multi-view UAV images during training.Meanwhile,by designing a multi-scale discriminator,the proposed method builds a cross view adversarial architecture of network for object detection,which improves the net-work ability of the feature alignment between different views.Furthermore,the method pro-poses a optimization strategy by discriminator,adversarial and object detection loss function to improve the effectiveness during training.(2)Aiming at solving the difficulty of object detection caused by the complex back-grounds of UAV remote sensing images,this thesis proposes an adaptive dense estimation network for complex background object detection.The method generates the density map ground truth by object detection annotations.Moreover,the method designs a density adaptive estimation network and builds an adaptive dense estimation architecture of network for ob-ject detection in UAV remote sensing images with complex backgrounds.With the proposed architecture,this method not only detects objects,but also generates density adaptive estima-tion map which takes the object center point position as the focus and promotes the learning ability of the entire feature extraction network on the object distribution.Furthermore,the method proposes a optimization strategy by density estimation loss function to improve the effectiveness during training.(3)Aiming at improving the poor object detection results caused by multi-scale objects in UAV remote sensing images,this thesis proposes a global density fused network for multi-scale object detection.This method proposes global density network,which increases the perception field of view of the multi-scale feature extraction part in the network and improves the network’s ability of learning global information with objects at different scales.Mean-while,the method builds a global density fused architecture of network for multi-scale object detection.The proposed method fuses the designed global density network to general object detection networks,which facilitates the entitle network to learn the distribution of objects with different scales.Furthermore,the method proposes a optimization strategy for object regression network to improve the effectiveness during training.Aiming at solving the difficulties of object detection due to variable viewing angles,com-plex backgrounds and multi-scale objects in UAV remote sensing images,this thesis proposes three object detection approaches.Contributions to the thesis include:(1)proposes a cross view adversarial network for object detection in UAV remote sensing images,which intro-duces single-view satellite images and builds a cross view adversarial architecture to improve the network ability of the feature extraction and object detection between different views.(2)proposes an adaptive dense estimation network for complex background object detection,which generates the density map ground truth by object detection annotations and designs an adaptive dense estimation architecture with optimization strategy to promotes the learning ability of the network on the object distribution and object detection.(3)proposes a global density fused network for multi-scale object detection,which designs global density network with optimization strategy to promotes the learning ability of the network on global feature learning and multi-scale object detection.The experiments in three approaches are conducted on two challenging public UAV benchmark datasets for qualitative and quantitative evaluation.The experimental results show that the proposed approaches are effective and superior.Furthermore,three datasets from real projects with variable perspectives,complex backgrounds and multi-scale target scenes are selected as additional datasets respectively in the experiments of the three approaches.And the experimental results are successfully applied to these projects.