Performance Comparison of Various YOLO Architectures on Object Detection of UAV Images
Teddy Surya Gunawan, Islam Mohamed Mahmoud Ismail, Mira Kartiwi, Nanang Ismail
Abstract
Today, the rapid development of deep learning offers an extraordinary opportunity to enhance the performance and efficiency of various industries, including business, the military, medicine, and transportation. Using deep learning algorithms in the transportation industry, for instance, makes UAVs vital and efficient in this industry. Current Unmanned Aerial Vehicles (UAVs) applications in transportation systems encourage the development of object detection methods to collect real-time traffic data using UAVs. Due to the versatility and portability of UAVs, particularly drones, individuals require systems that operate with UAVs to identify objects in real-time for military, safety observation, and protection. The culmination of the evolution of computer vision technology is the development of sophisticated algorithms centered on extensive training and testing datasets. This research aims to compare the performance of object detection of UAV images using various YOLO architectures. Tiny YOLOv3 and YOLOv5s models were implemented to extract the object’s features and classify them into the dataset’s multiple classes. This paper selected the VisDrone2019 dataset for its various object classes: pedestrian, person, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor. Results demonstrated that YOLOv5s have acceptable precision and processing speed.