Oriented Object Detector In Aerial Images Based On YOLOv5

With the widespread popularity of UAVs in application fields such as homeland security,urban security,etc.,UAVs and computer vision technology are getting more and more connected.Moreover,object detection is a basic task of computer vision,its performance has an extremely important impact on the performance of the entire system.Due to the orientation of the object in UAVs’image,the traditional horizontal bounding boxes is not satisfactory in this scene,so a large number of oriented object detector has been proposed.But the existing oriented object detector runs at the fastest speed of only25 fps on the 2080Ti,which is too costly to deploy for the real-time detection requirements of UAVs.In addition,the commonly used loss functions in the field of oriented object detection cannot accurately measure the gap between the prediction and ground truth,resulting in the contradiction,which is the optimization direction of reducing the loss is not necessarily the optimization direction of improving accuracy,this contradiction finally reduce the training efficiency of the network.In order to improve the above shortcomings,my thesis will improve the regression loss about rotated rectangular boxes,and propose a novel oriented object detector that balances accuracy,model size,and speed.In view of the real-time detection requirements of the algorithm,the scale variability and arbitrary orientation of the aerial image object,my thesis constructs a better YOLOv5 as the baseline and then proposes a oriented object detector named R-YOLOv5.Aiming at the problem of boundary discontinuity in the field of oriected object detection,my thesis also proposes a novel regression loss function named R-CIo U Loss,which measures the gap between the prediction and ground truth by considering RIo U,the distance of center points,and the shape parameters.Compared with other regression loss,R-CIo U Loss can more accurately describe the intersection state between two rotated rectangular boxes,and can effectively avoid the inconsistency between the loss and the metric.R-CIo U Loss can also improve the optimization efficiency and accuracy of the oriented object detector.Finally,in order to reduce the the deployment cost of R-YOLOv5,my thesis compresses the R-YOLOv5 by integrating the convolutional layer and the BN layer,using smaller parameter network structure.Experiments shows that R-YOLOv5 can achieve excellent detection performance on deployment platforms with different computing power.Among them,the smallest R-YOLOv5n achieves 73.26 m AP₅₀and65.79 fps on DOTAv1-test with only 2.02 M parameters,4.8 MB model memory and12.6 GFLOPs.Compared with baseline,ablation experiments show that R-YOLOv5+R-CIo U Loss can maintain accuracy while reducing the amount of parameters by 90%,reducing memory of model by 88%,reducing the FLOPs by 89.8%,and increasing the fps by45.1%.R-YOLOv5+R-CIo U Loss is the algorithm with the best comprehensive performance in terms of detection accuracy,model size and speed in the field of oriented object detection,and has good generalization in UAV aerial imagery scenes.