Performance Comparison of YOLO Variants for Object Detection in Drone-Based Imagery

Abstract

With the advances and development in technology, drones mounted with cameras are being used in various real-life applications like surveillance, vehicle tracking, agriculture, disaster management, pedestrian monitoring, search and rescue operations, rapid delivery services and aerial photography. Hence, there is a high demand for understanding the visual data collected through these drones for their efficient use in these areas. Computer vision is an obvious solution for the detection and tracking of objects in drone-captured images and sequences. However, the already developed techniques and algorithms cannot be directly applied to the drone platforms because of the inherent challenges in this field and the requirement of balance between accuracy and speed due to constraints of limited onboard computation power and memory of the embedded devices being used in drone systems. Thus, there is a requirement of lightweight real-time object detectors for drone systems. Out of the various real-time detectors, YOLO series models have emerged as the fastest object detection algorithm, which makes them a potential contender for drone-based object detectors. In this paper, we have compared the performance of the various YOLO algorithms, namely, YOLOv3, YOLOv4 and YOLOv5 for multiclass object detection in drone-based images. These algorithms have been evaluated and compared using the VisDrone2019-Det benchmark dataset.