Research On Target Detection Algorithm Of UAV Based On Deep Learning

The development of the unmanned aerial vehicle(UAV)industry was explosive in 2019.All kinds of UAV products are emerging in an endless stream.UAV has great application value and research prospect in military,industry,entertainment and other fields.It is the same important research problem as the development of UAV to be able to accurately detect such targets.The characteristics of UAV targets are small size,relatively complicated flight environment,and the flight speed and attitude are easy to change.In general,UAV targets have weaker visual characteristics.For these reasons,traditional detection algorithms are not effective at detecting such targets.Therefore,this treatise focused on several factors that affect the detection performance of the detected model,including algorithm model,scale changes,attitude changes,background complexity changes,occluded area changes and quantity changes.According to the conclusion,a detection model for UAV has been constructed to meet the purpose of detecting UAV targets under different task scenarios.Based on the above research background and needs,this treatise has studied the following content.(1)Mainstream detection algorithm have been researched.Focused on analyzing and comparing the connection and difference between R-CNN series algorithm,YOLO series algorithm and SSD algorithm,specifically including the detection principle,network framework,implementation steps and improvements.At the same time,the performance of Faster R-CNN,YOLOv3,SSD for detecting UAV targets under insufficient training data were studied.(2)Taking SSD algorithm as the main research object,the impact of various influencing factors on detection performance have been in-depth studied.The results show that the UAV targets have a high degree of symmetry,and the attitude changes have little effect on the detection performance.In the case of small background interference,the model’s limit detection ability for small UAV targets are about 21×16 pixels.The smaller the target size,the more sensitive it is to changes in background contrast and occlusion area.In general,when the occlusion area did not exceed 50%,it could be detected by the model.(3)For situations where there are multiple targets,two conjectures were put forward for the problems of serious missed detection and experimental verification was carried out.Starting with the training data,the model’s ability to detect “many and small” UAV targets has been initially improved.(4)Marked 1500 UAV target images as training samples,containing different scales,attitudes and backgrounds.200 images of multi-target with sky as background also have been marked as training data.A comprehensive test set of SUA targets has been established,containing 150 images of similar targets with different backgrounds.According the researching for the statistical research on the characteristics of UAV target data and the research conclusions,a special UAV target detection model has been established.Through model improvement and data enhancement,the detection model’s ability has been further improved and the accuracy has reached more than 90%.