Aerial Object Detection Using Different Models of YOLO Architecture: A Comparative Study

Abstract

AbstractAerial object detection is a key to many functionalities like security systems, pedestrian counting, animal population estimation, security surveillance and many more. Previously used traditional machine learning approaches made use of handmade features and algorithms that failed to generalise on a larger data set. Therefore, deep learning models have outperformed the traditional machine learning models, especially in the computer vision field. Aerial object detection is a subdomain of object detection that has been a hot topic of interest in recent years. Transfer learning has emerged as one of the go-to methods to adapt models well on a small data set. This paper proposes a comparative study of the performance of three state-of-the-art object detection models on an aerial images data set using Transfer learning. The paper aims at studying how well a model can adapt to aerial images using transfer learning. The three models used are the YOLOv5m6, YOLOv5x6 and the YOLOv5l6 models based on the YOLO architecture. The data set used is the Vehicle Detection in Aerial Imagery (VEDAI) data set. The study finds out that the YOLOv5m6 model outperforms the other two models.KeywordsMachine learningDeep learningComputer visionObject detectionAerial object detectionTransfer learningYOLO