Fruit counting using autonomous flight of small UAV

Abstract

A fundamental study of autonomous small unmanned aerial vehicle (UAV) flights for application in agricultural fields is conducted in this study. Although some types of UAVs can trail predetermined routes based on GPS information, autonomous flights are difficult to accomplish in small and low-altitude regions. In this study, a UAV is connected to a PC through Wi-Fi to pilot the UAV. The small UAV captures images and then sends them to the PC. The number of fruits in each image is determined based on the recorded images. For fruit detection, a deep-learning method and an image-processing method are compared. Results show that citrus fruits can be detected using both methods, and that the image processing method yields more accurate counting. Total of 151 fruits were over-detected out of 139 using our proposed method using image processing technique. In the result, 14 fruits were counted twice, and some fruits could not detected. The F1 value of the detection was over 0.9, meaning accurate fruit counting via automatic flight of small UAV could be done. However, some deep-learning methods such as YOLOv3 fail to detect some fruits, thereby resulting in an underestimated number of fruits. The YOLOv3 detected only 95 out of the total 139 fruits. Although precision was over 0.9, recall value was under 0.7. Our results suggest that the autonomous flight of small UAVs can afford more efficient as well as laborless monitoring in agricultural fields and houses.