Research On Unmanned Aerial Vehicle Detection Based On Deep Learning

The widespread use of Unmanned Aerial Vehicles(UAV)has caused hidden dangers to air safety.The monitoring of UAV has become a problem that must be solved at present.UAV detection based on image and video signals has become one of the solutions to the problem.In this thesis,we focus on the UAV detection algorithm in a single image.Due to the large number of UAV models,in the UAV detection mission,the target has a large scale range in the image,and the background of the UAV is complex and variable under different observation angles,so the stable performance of target detection is facing a series of challenges.In view of the above problems,this thesis chooses the detection algorithm based on deep learning as a technical solution,and carries out related exploration and research.This thesis first collects and annotates a UAV dataset.The dataset contains 7,229 images and 9,322 UAV targets,covers a variety of UAV models,different target sizes,different flight conditions and a variety of scenes,and includes motion blur,lighting changes and other common interference factors in the detection process.In order to solve the problem of small-scale UAV detection,this thesis proposes an improvement anchor method for small-scale UAV targets.Through anchor template learning and anchor compensation strategy,the performance of Faster-RCNN algorithm is improved for small-scale UAVs.Firstly,we use the clustering method to learn the anchor aspect ratio and anchor size suitable for the UAV dataset,so the anchor template is more targeted to the UAV target.Secondly,compensation strategy improves the number of anchors that is matched to a target,at the same time,it alleviate the problem that small-scale targets are difficult to match to the anchor.Experimental results show that the new algorithm improves the performance of small-scale UAV detection.In theory,the algorithm can also be used in the detection of other specific types of small-scale targets.In order to solve the problem of UAV detection in complex background,this thesis proposes a detection algorithm based on I&M-Heads Boosting.The I&M-Head uses the feature extractor based on Inception network to improve the robustness of the feature,and refines the background category by designing a multi-background classification framework of Maxout structure,thus improving the performance of classification under complex background.Finally,the author designed a UAV target detection algorithm based on I&M-Heads Boosting,which uses the idea of boosting method to enhance the classifier's ability.Experiments show that the new algorithm significantly improves the performance of UAV detection in complex backgrounds.In theory,the algorithm can also be used in the detection of other types of targets in complex backgrounds.Finally,the author fuses the network framework of the two detection algorithms proposed in this thesis.The experimental results verify that the performance of the fusion detection algorithm is superior to some advanced detection algorithms.