Improving the performance of YOLO-based detection algorithms for small object detection in UAV-taken images

Recent advances in computer vision yield emerging novel applications for cameraequipped unmanned aerial vehicles such as object detection. The accuracy of the existing object detection solutions running on images acquired by Unmanned Aerial Vehicles (UAVs) is limited when compared to the performance of the object detection solutions designed for ground-taken images. Existing object detection solutions demonstrate lower performance on aerial datasets because of the reasons originating from the nature of the UAVs. These reasons can be summarized as: (i) the lack of large drone datasets with different types of objects, (ii) the larger variance in both scale and orientation of objects in drone images, and (iii) the difference in shape and texture of the features between the ground and the aerial images. Due to these reasons, YOLO-based models, a popular family of one-stage object detectors, perform insufficiently in UAV-based applications. In this thesis, two improved YOLO models: YOLODrone and YOLODrone+ are introduced for detecting objects in drone images. The performance of the models are tested on VisDrone2019 and SkyDataV1 datasets and improved results are reported when compared to the original YOLOv3 and YOLOv5 models.