Research On Object Detection Algorithm Of UAV Aerial Images In Complex Scenes

With the rapid development of unmanned aerial vehicle(UAV)technology,there are more and more applications of UAVs in many fields.For example,in the electrical field,circuit inspection is completed by UAV,which greatly improves the accuracy and efficiency of inspection.Therefore,the object detection technology of UAV aerial images has attracted the attention and research of many scholars.However,the background of UAV aerial images is complex and changeable,the object scale changes drastically,and there are a large number of small objects that occlude each other,which brings great challenges to object detection.In addition,how to deploy a model with a large amount of calculation and parameter real-time object detection on UAVs has also become one of the current challenges.Based on the above background and the difficulties existing in the object detection algorithm of UAV aerial images,this paper proposes a series of object detection algorithms for UAV aerial images for different application scenarios.The main innovations and research work are as follows:(1)An object detection algorithm fuses attention mechanism and context for UAV images Eagle-YOLO is proposed.Aiming at the difficulty of large numbers of small objects and complex and changeable backgrounds in UAV aerial images,this paper integrates the LKA attention mechanism on the basis of YOLOv5 to make the model focus on the object area to be detected to obtain more detailed information of the object.And combining the multi-scale feature fusion network of learnable weight parameters to fuse context information,so as to better improve the detection performance of small objects.In order to further improve the performance of the model,this paper proposes the Eagle-Io U loss,which effectively solves the problem of unstable gradient and slow convergence of the original model during the previous training.Finally verifies it on the Vis Drone dataset and the results show that compared with the baseline model,the proposed model improves m AP and AP50 by 2.86% and 4.23%,respectively.(2)An object detection algorithm for complex UAV images based on rotating frame recognition Rotate-YOLOv5 is proposed.Aiming at the problem that the horizontal frame cannot accurately locate the object in the scene where the objects are closely arranged and the object size is slender,firstly,YOLOv5 is used as the baseline model,and the calculation of the angle loss is added to the original loss function to apply the object detection of the rotating frame.The original detection heads are decoupled to predict different tasks separately to accelerate the convergence speed of the model.Then,the angular regression task is converted into a fine-grained classification task by means of densely encoded labels,which solves the problem that the loss value suddenly increases at the boundary caused by the periodicity of the angle.At the same time,this paper also uses the KFIo U loss to improve the computational efficiency of the model,and finally verified on the DOTA dataset.The results show that the detection accuracy m AP of RotateYOLOv5 proposed in this paper reaches 73.61%.(3)A lightweight object detection algorithm for UAV images YOLOv5-Light is proposed.Aiming at the problem that the amount of model parameters and calculations are too large to be deployed on UAVs for object detection.In this paper,YOLOv5 is used as the basic network for lightweight design of its model structure.Firstly,the C3 module and Conv module with large amounts of original parameters and calculation are optimized into C3 Ghost module and GSConv module.Then,a parameter-free SPD module is used to reduce the problem of object fine-grained information loss,thereby improving the detection accuracy of small objects.Finally,the trained model is compressed by means of knowledge distillation,which further reduces the amount of computation and parameters of the model,making the model suitable for deployment on edge devices.